2018 年 4 月的一个早晨，春日的阳光照进哥伦比亚大学舍默霍恩大厅一间研讨室的东窗，我（威金斯）走到黑板前，开始解释量化物化，即数值对应经验观察成为事物的神奇过程。我带着阿道夫·凯特勒的故事在黑板上画了一条永恒的“正态曲线”。阿道夫·凯特勒试图利用他获得的苏格兰士兵体格测量数据来揭示理想的人。数学家称其为高斯曲线，智商测试中臭名昭著的“钟形曲线”也备受争议，正态曲线对自然科学家来说意味着数据揭示了一些真实的东西，甚至是一些超越的东西。我转向学生们，希望从他们的眼睛里看到他们和我一样兴奋。一个学生迎上我的目光，把手掌朝天问道：“我们现在可以谈谈 Facebook 吗？”
One April morning in 2018, as the spring sunlight streamed into the eastern window of a seminar room in Schermerhorn Hall at Columbia University, I (Wiggins) went to the chalkboard to explain quantitative reification, the magical process whereby a numerical correspondence to empirical observation becomes a thing. Armed with the story of Adolphe Quetelet, who aimed to reveal the ideal man using data he obtained about the physical measurements of Scottish soldiers, I traced on the blackboard the immortal “normal curve.” Known to mathematicians as the Gaussian curve, and contested as the notorious “bell curve” of IQ tests, the normal curve signifies to natural scientists that data has revealed something real, even something transcendent. I turned to the students, hoping to see in their eyes that they shared my excitement. One met my gaze and, turning his palms to the heavens, asked: “Can we talk about Face-book now?”
那天早上，报纸和数字新闻都预示着华盛顿即将燃起熊熊烈火，一切隐瞒都将烟消云散。硅谷一家改变文化的科技公司的首席执行官被传唤到美国参议院。《纽约时报》解释说，参议员们代表全体公民，试图了解数百万人的个人数据是如何被泄露的，包括我们这样的学生，这些数据被用来不良后果违反了我们的隐私和政治程序规范。1在本周的国会证词结束时，学生们认识到了我们的民选官员对数字媒介现实的理解与他们从小在算法的熏陶下获得的个人知识之间存在的差距。
That morning, newspapers and digital news feeds alike heralded a hot fire about to burn in Washington, one that would melt down all concealment. The irreverent CEO of a culture-changing tech company in Silicon Valley was being called before the United States Senate. On behalf of all citizens, the senators sought to understand how the personal data of millions of people, including students such as ours, was compromised, The New York Times explained, used for ill ends that violated our norms about privacy and our political process.1 By the end of the week’s congressional testimony, students recognized the extent of the gap between how our elected officials understood their digitally mediated reality and their personal knowledge from growing up with algorithms.
数据的故事充满了竞争：定义什么是真实的竞争，使用数据来提升自己权力的竞争，有时，还有使用算法和数据照亮黑暗、赋予无助者权力的竞争。这本书的创作源于我们对数百名好奇学生的教学，以及我们作为科学史学家和执业数据科学家的亲身经历，以及作为公民试图理解我们如何生活在这个算法介导的现实中，以及我们如何选择不同的生活方式。像所有用户、开发人员和技术主体一样，我们试图理解这一切的发展方向，以及我们将如何共同塑造未来。我们试图讲述一个关于思想和技术的故事，同时也讲述一个关于真理和权力的历史。
The story of data is replete with contests: contests to define what is true, contests to use data to advance one’s power, and, on occasion, contests to use algorithms and data to shine a light into darkness and to empower the defenseless. This book grew out of our teaching hundreds of inquisitive students, along with our own experiences, as a historian of science and as a practicing data scientist, and as citizens trying to understand how we came to live in this algorithmically mediated reality and how we might choose to live differently. Like all users, developers, and subjects of technology, we are trying to make sense of where it is all headed as well as how we collectively will shape that future. We’ve attempted to tell a story of ideas and technologies but also a history of truth and power.
放下粉笔，我们一致认为凯特勒将大放异彩。但首先，我们需要解释一下这位名不见经传的比利时天文学家是如何融入数据故事的：数据及其分析手段如何从国家关注的领域转移到大学、军队和私营企业。
Putting down the chalk, we agreed that Quetelet would have his day. But first, we would need to explain how an obscure Belgian astronomer fits in with the story of data: how data and the means for analyzing it would move from a concern of the state to universities, the military, and private corporations.
我们在这里使用“数据”来指代几乎无处不在的数据驱动算法决策系统。我们探索数据是如何产生和整理的，以及为处理这些数据而开发的新数学和计算技术如何塑造人、思想、社会、军事行动和经济。数据带来了权力，包括塑造人们认为是真实的事物的权力。尽管技术和数学数据的故事本质上涉及国家、企业和个人之间不稳定的博弈。
We use “data” here as shorthand for the expanse of data-driven algorithmic decision-making systems surrounding us nearly everywhere. We explore how data was created and curated as well as how new mathematical and computational techniques developed to contend with that data serve to shape people, ideas, society, military operations, and economies. Along with data comes power, including the power to shape what is perceived to be true. Although technology and mathematics are at its heart, the story of data ultimately concerns an unstable game among states, corporations, and people.
因此，那天早上，我们不仅谈论数据，还谈论数据主导的世界的利害关系。
And so, on that morning, we spoke not just of data, but of the stakes for a world mediated by data.
背景
Background
开设数据如何发生的课程的想法诞生于 2015 年 11 月，当时我们正与几名来自工程和人文背景的哥伦比亚大学本科生共进晚餐。当时，我们推测学生们对数据科学的历史非常感兴趣，我们认为我们结合互补的观点将提供一个有用的视角，为工程师和非技术人员提供新材料。当我们在 2017 年 1 月第一次教授这门课时，我们很快意识到，学生们不仅对我们如何走到今天感兴趣，而且还在寻找一个分析和可操作的框架来理解数据的伦理和政治。2
The idea for a class on how data happened was born in November 2015, at a small dinner conversation with a few Columbia undergraduates drawn from a mix of engineering and humanities backgrounds. At the time, we conjectured students were very interested in the history of data science, and we thought that our combined, complementary perspectives would give a useful view, with material new both to the engineers and to the non-technologists alike. When we taught the class for the first time in January 2017, we quickly realized that the students were not only interested in how we got here but were searching for an analytic and actionable framework for understanding the ethics and politics of data.2
我们所说的“政治”并非狭义的“投票”，而是“与权力动态有关或与之相关的政治”。我们的目标是提供一个框架，以理解数据在重新调整权力（企业权力、国家权力和人民权力）方面所发挥的持久作用。我们的历史弧线提供了关键的杠杆，引导我们共同理解现在以及我们可以用来塑造未来的武器和工具。
By “politics” we don’t mean a narrow definition as “voting” but “of or relating to the dynamics of power.” Our goal is to provide a framework for understanding the persistent role of data in rearranging power: corporate power, state power, and people power. Our historical arc offers critical leverage that points us toward a shared understanding of the present as well as the weapons and tools at our disposal for shaping the future.
本书
This Book
每段历史都有开始的地方，我们发现一个有用的起点是十八世纪末，大约是“统计”一词首次进入英语的时候。我们的故事围绕着收集统计数据的艰苦工作展开。数据，包括建立收集和公开数据的基础设施，以及开发用于研究数据的新数学和计算技术——理解和主张数据的新方法，以及使用这些主张进行决策，往往会深刻地改变生活，无论是好是坏。在每一章中，我们都将考虑一种思想转变。我们讨论如何开发一种新的技术和科学能力；谁支持、推进或资助了这种能力或转变；这种转变是如何受到质疑的；以及这种新能力如何重新安排权力——改变谁可以做什么、从什么做起以及对谁做。3我们不仅关注军事或金融权力的重新安排，而且更普遍地关注具有伦理和政治价态的转变：即数据影响权利、重新安排危害或支持（或阻碍）正义的转变。
Every history must begin somewhere, and we found a useful starting point to be the end of the eighteenth century, around the time the word “statistics” first entered the English language. Our story tacks between the hard work of collecting data, including building infrastructure to collect and make it public, and the development of new mathematical and computation techniques for studying data—new ways of understanding and making claims about that data and using those claims to make decisions, often profoundly changing lives, for better and for worse. In each chapter we consider one intellectual transition. We discuss how a new technical and scientific capability was developed; who supported, advanced, or funded this capability or transition; how this transition was contested; and how this new capability rearranged power— changing who could do what, from what, and to whom.3 We focus not only on rearrangements of military or financial power, but more generally on those transitions with an ethical and political valence: those in which data affects rights, rearranges harms, or supports—or thwarts—justice.
《数据是如何产生的》以数据为治国服务为开篇，随后转向数据改善社会，以及数据接受数学洗礼，从而创建了一个名为“数理统计”的新学术领域。第二部分以数据在第二次世界大战中用于密码破译的军事应用开篇，这与数字计算的诞生相吻合。我们从英国的布莱切利园到美国的贝尔实验室，再到二战后数据的商业和工程应用。从企业权力过渡到国家权力和“人民权力”的反应，我们探索了数字化个人记录保存对我们对隐私理解的影响，特别是 1970 年代公众希望通过隐私来防止国家权力过度扩张。我们追溯了“人工智能”领域的首次诞生和消亡及其复兴，它以“机器学习”的形式从灰烬中重生，而机器学习基于不断增长的关于公民、消费者和军事对手的数据存储库。
How Data Happened begins with data in the service of statecraft, before turning to the usage of data to improve society and the mathematical baptism of data with the creation of a new academic field called “mathematical statistics.” The second part opens with the martial application of data in World War II for codebreaking, coinciding with the birth of digital computation. We follow the thread from Bletchley Park in England to Bell Labs in the United States, and to the business and engineering applications of data in the wake of World War II. Transitioning from corporate power to the reactions in state power and in “people power,” we explore the impact of digital, personal record keeping on our understanding of privacy, particularly the public desire for privacy as a defense from overreaching state power in the 1970s. We trace the first birth and death of the field of “artificial intelligence” and its renaissance, rising from the ashes in the form of “machine learning” based on ever growing repositories of data about citizens, consumers, and military adversaries alike.
本书的最后一部分将过去与我们的现在和未来联系起来。我们通过研究财务安排和商业模式，讨论了数据和权力如何从国家关注转向企业关注，这些财务安排和商业模式使得单个公司能够借助数据赋能技术迅速主导整个行业。关于道德的争论已经提出了许多潜在的企业权力补救措施；我们追溯了应用伦理在研究中的历史，以及它如何影响数据赋能算法作为产品的部署方式，塑造我们的个人和政治现实。
The last section of the book connects this past to our present and future. We discuss how data and power moved from a state concern to a corporate concern, by looking to the financial arrangements and business models that have allowed single corporations to dominate entire sectors rapidly with the help of data-empowered technologies. A contested debate over ethics has framed many of the potential remedies to corporate power; we trace the history of applied ethics in research and how it has impacted the way that data-empowered algorithms are deployed as products, shaping our personal and political reality.
最后，我们来讨论一下未来。无论做出预测有多困难，组织我们对未来的理解的一个精辟方法是描述当前权力之间的竞争以及这些竞争将在哪些领域产生结果。在本书的最后，我们将探讨我们认为目前企业权力、国家权力和人民权力之间最重要的竞争，以及新形式团结的可能性。这些竞争的解决将塑造我们的集体未来，更倾向于正义——或者可能不是。
Finally, we discuss the future. However difficult it may be to make predictions, one incisive way to organize our understanding of the future is to describe the present contests among powers along with the arenas in which these contests will be decided. We close the book by looking at what we consider to be the most important present contests among corporate power, state power, and people power, along with the possibility of new forms of solidarity. The resolutions of these contests will shape our collective future, leaning more toward justice—or perhaps not.
我们的目标是对历史有一个切实可行的理解。我们不会回避自己作为公民、技术专家和个人的角色；我们是这些产品的用户，而且——正如早在 20 世纪 70 年代就指出的那样，由于我们处在广告经济中，因此我们也是产品。
Our goal here is an actionable understanding of history. We will not shy away from our own roles as citizens, technologists, and individuals; we are users of these products and— as noted as early as the 1970s, since we are in an advertising economy, therefore we are also the product.
我们在书中提出了两种互补的观点，每种观点都有局限性和偏见。威金斯在哥伦比亚大学任教二十多年，一直在开发机器学习方法来理解生物学和健康，自 2013 年以来，他担任《纽约时报》首席数据科学家，开发和部署机器学习方法和产品。在 CP Snow 的“两种文化”的另一边，琼斯是一位科学史学家，他追踪了数学方法如何自 17 世纪的“科学革命”以来，思考和辩论成为研究自然和政治的重要权威方式。特别是在研究数据的多少种用途会放大差距时，我们大量借鉴了许多揭露这些过程的学者和活动家的启发性著作。许多（如果不是大多数的话）最尖锐和最有才华的批评家来自与我们截然不同的背景和经历——两位终身任职的白人男性学者。我们的工作始终建立在他们的劳动和洞察力之上并加以赞扬。我们将指出关于数据赋能算法和技术的全球影响的优秀文献——以及数据在我们的社会、经济和教育机构组织中的历史。我们更现代的材料主要集中在美国。我们提供了尾注，不仅反映了在哪里可以找到更多关于我们在课堂上涵盖和在学术出版物中撰写的主题，还反映了许多重要的作品，包括学术文献，我们鼓励读者参与以更深入地理解。
We bring to the book two complementary perspectives, each with limitations and biases. Wiggins has been developing machine learning methods for understanding biology and health for over twenty years as faculty at Columbia and, since 2013, developing and deploying machine learning methods and products as chief data scientist at The New York Times. On the other side of C. P. Snow’s “two cultures,” Jones is a historian of science who has tracked how mathematical ways of thinking and arguing became a crucial authoritative way to study nature and politics from the “Scientific Revolution” of the seventeenth century forward. Particularly when examining how many uses of data amplify disparity, we draw heavily upon the illuminating writing of the many scholars and activists who have exposed these processes. Many, if not most, of the most trenchant and luminous critics crucially come from backgrounds and have experiences quite different from ours— two tenured white male academics. Our work builds upon and celebrates their labor and insight throughout. We will point to excellent literature on the global impact of data-empowered algorithms and technologies—and to the histories of data in the organization of our societies, economies, and educational institutions. Our more contemporary material focuses primarily on the United States. We have provided endnotes to reflect not only where to find out more about the topics we cover in class and write about in our scholarly publications, but many important works, including the scholarly literature, that we encourage readers to engage for deeper understanding.
我们力求清晰描绘出企业权力、国家权力和人民权力之间历史和当前的紧张关系，重点关注数据在确定真相和塑造这些权力之间的竞争中所起的作用。我们希望展示我们如何共同走到这一步，说明那些小小的巧合、主观的设计选择和欺骗，它们僵化成了看似“必须如此”的事情。了解这些转变和偶然性将揭示过去如何解决类似的问题。这反过来又有助于我们想象如何打破和重置系统骨架，这些系统有时会赋予无助者权力，但更多时候却强化了有权势的人。
We seek to give a clear picture of historical as well as current tensions among corporate power, state power, and people power, focusing on the role of data in establishing truth and shaping the contests among these powers. We hope to show how we collectively got here, to illustrate the small coincidences, subjective design choices, and deceptions that ossified into what only seems like things that “must be that way.” Understanding these transitions and contingencies will reveal how similar problems were solved in the past. This will, in turn, help us picture how we could break and reset the bones of systems that sometimes empower the defenseless— yet have more often strengthened the empowered.
通过展示看似不可改变的结果如何取决于过去的选择，我们可以看到如何共同选择不同的未来。
By showing how apparently immutable results hinged on past choices, we can see how we can collectively choose a different future.
Technology is neither good nor bad; nor is it neutral.
–克兰兹伯格第一技术定律，1986 年
–Kranzberg’s first law of technology, 1986
我在密歇根大学教授一门课程，名为“互联网是一场垃圾大火”，我不需要向任何人解释这是什么意思……我们忍受这种情况很长时间了；我们似乎不知道有什么不同。
I teach a course at the University of Michigan called “the Internet is a trash fire,” and I don’t have to explain to anybody what that means. . . . We put up with this for a long time; we don’t seem to know anything different.
–Lisa Nakamura，2019 年
–Lisa Nakamura, 2019
2014 年 12月，计算机科学家 Hanna Wallach 在蒙特利尔会议中心向一群技术专家、律师和活动家呼吁进行一场革命。在与研究“机器学习”的顶尖计算机科学家交谈时，她提出，她所在的领域迫切需要审视他们正在开发的算法以及算法所支持的技术如何挑战我们的“公平、问责和透明”价值观。尽管哲学家、社会学家和律师多年来一直在敲响警钟，但在这里，一位技术界的指定成员，在微软拥有令人垂涎的研究职位，借鉴了这项重要工作，并呼吁同事们改进他们的研究，正是通过认识到他们的算法系统需要公平和问责，才能更好地开展工作。
In December 2014 at the Palais des Congrès de Montréal, the computer scientist Hanna Wallach advocated for a revolution before an audience of technologists, lawyers, and activists. Speaking to top computer scientists working on “machine learning,” she proposed that her own field desperately needed to interrogate how the algorithms they were developing, and the technologies the algorithms empowered, challenged our values of “fairness, accountability, and transparency.” While philosophers, sociologists, and lawyers had been sounding the alarm for years, here, an anointed member of the technical community, with a coveted research position at Microsoft, drew on that critical work and called for colleagues to improve their research, to do better work precisely by recognizing the need for fairness and accountability from their algorithmic systems.
瓦拉赫的演讲绝不是一场旷日持久的抗议，而是在应用机器学习领域最重要的会议上，有人将论文直接贴在大教堂的大门上。瓦拉赫诊断出了问题——一个超出计算机科学传统学科范围的问题。她承认，问题的解决方案不会来自计算机科学内部，而是要求与其他领域的人合作。“很少有计算机科学家或工程师，”瓦拉赫解释说，“会考虑开发用于分析天文学数据的模型或工具，而无需天文学家的参与。那么，为什么这么多分析社会数据的方法是在没有社会科学家参与的情况下开发的呢？” 1
Far from a protest in the wilderness, Wallach’s talk, delivered at the most important conference in applied machine learning, was a posting of theses directly on the door of the cathedral. Wallach diagnosed the problem—one outside the traditional disciplinary scope of computer science. Admitting that the solutions to the problem would not come from within computer science, she instead demanded collaboration with those from other fields. “Few computer scientists or engineers,” Wallach explained, “would consider developing models or tools for analyzing astronomy data without involving astronomers. So, why, then, are so many methods for analyzing social data developed without the involvement [of] social scientists?”1
Wallach 敦促人们更深入地认识到偏见潜入机器学习者创建的模型的方式，并警告说，仅仅因为数据集可用而研究数据集本身就存在风险。例如：虽然从 Twitter 用户那里获取和分析数据相对简单，但这些数据很难代表整个美国人口。她敦促研究人员“开始跳出机器学习社区通常接受的算法框架，转而关注开发和使用机器学习方法来分析有关社会的现实世界数据所涉及的机会、挑战和影响。” 2
Wallach urged a deeper recognition of the ways that biases creep into models created by machine learners and warned of the risks inherent in studying data sets simply because they were available. As an example: while it’s relatively straightforward to obtain and analyze data from Twitter users, these data are hardly representative of, say, the US population overall. She urged researchers “to start thinking outside the algorithmic boxes typically embraced by the machine learning community and instead focus on the opportunities, challenges, and implications involved [in] developing and using machine learning methods to analyze real-world data about society.”2
在瓦拉赫发表演讲时，分析现实世界的社会数据已经成为互联网巨头谷歌、Facebook 和亚马逊商业模式的核心，更不用说它也是美国、英国、以色列和中国情报机构的核心。可以说，这些公司和机构很少考虑瓦拉赫演讲中提到的公平性和责任感问题。核心问题不仅仅是学术问题，也不仅仅是研究界的问题转移焦点。
Analyzing real-world data about society was, by the time Wallach was speaking, already the heart of the business models of the internet giants Google, Facebook, and Amazon—not to mention central to the intelligence agencies of the United States, the United Kingdom, Israel, and China. Suffice to say, these corporations and agencies rarely incorporated these questions of fairness and accountability animating Wallach’s talk. The issues at heart were not simply academic nor merely a question of a research community shifting its focus.
2000 年以后，一群活跃的社会科学家逆着互联网乌托邦主义的潮流而动，对数据驱动的互联网在商业、教育和治理方面的表现表示担忧。在 2011 年牛津举行的研讨会上，达娜·博伊德和凯特·克劳福德认为：“大数据时代已经开始。”在包括“互联网发明者”之一温特·瑟夫等名人在内的观众面前，研究人员试图激发社会对大数据早期时代的批判性思考：
Swimming against the tide of internet utopianism after 2000, a lively group of social scientists had signaled concerns about the data-driven internet in commerce, education, and governance. At a 2011 symposium held at Oxford, danah boyd and Kate Crawford argued, “The era of Big Data has begun.” In front of an audience including such luminaries as Vint Cerf, one of the “inventors of the internet,” the researchers sought to provoke the community to think more critically about the incipient age of big data:
大规模搜索数据会帮助我们创造更好的工具、服务和公共产品吗？还是会引发新一轮侵犯隐私和侵入性营销浪潮？数据分析会帮助我们了解在线社区和政治运动吗？还是会被用来追踪抗议者和压制言论？大量数据会改变我们研究人类交流和文化的方式，还是会缩小研究范围并改变“研究”的含义？3
Will large-scale search data help us create better tools, services, and public goods? Or will it usher in a new wave of privacy incursions and invasive marketing? Will data analytics help us understand online communities and political movements? Or will analytics be used to track protesters and suppress speech? Will large quantities of data transform how we study human communication and culture, or narrow the palette of research options and alter what “research” means?3
研究人员借鉴奥斯卡·甘迪 (Oscar Gandy Jr.)、温迪·陈 (Wendy Chun) 和海伦·尼森鲍姆 (Helen Nissenbaum) 等早期批评人士的观点，开始记录企业和政府未能面对这些棘手问题对现实世界的影响，并呼吁进行重大变革。4我们不必声称已经充分掌握了这一庞大的研究成果，但可以举几个关键的例子。
Drawing upon earlier critical voices like Oscar Gandy Jr., Wendy Chun, and Helen Nissenbaum, researchers began documenting the real-world effects of corporations and governments failing to face these troubling questions—and called for dramatic change.4 Without claiming to capture adequately this enormous body of work, let us mention a few key examples.
2013 年，时任伊利诺伊大学香槟分校教授、现为麦克阿瑟“天才”奖学金获得者的萨菲亚·诺布尔发表了一篇批评谷歌搜索的文章。诺布尔写道：“在提供关于女性和有色人种的可靠、可信和历史背景信息方面，商业搜索表现不佳。”，尤其是黑人妇女和女孩。”表面上没有偏见的技术迅速复制并强化了针对黑人女性的种族主义和性别歧视。她写道：“继续研究这些现象，是一个质疑技术所谓中立性的机会，同时为社会正义和网上公平代表创造新的机会。” 5 2016 年，数学家 Cathy O'Neil 描述了自己的历程，从学者到华尔街员工，再到不受约束的数据和算法的批评者。在她的《数学毁灭武器》一书中，她探讨了数据科学的激励措施如何破坏其受试者的人性：“人们倾向于用数据轨迹取代人，将他们变成更有效的购物者、选民或工人，以优化某些目标。当成功以匿名分数的形式出现，而受影响的人仍然像屏幕上跳动的数字一样抽象时，这很容易做到，也很容易证明是合理的。” 6如果不改变这些激励机制，数据科学尽管前景光明，但会极大地改变一个又一个组织的目标，从大学到医学，再到社会福利，而牺牲的，主要是社会中最弱势的成员。 在这些关键诊断的同时，爱德华·斯诺登 (Edward Snowden) 揭露了美国及其“五眼”盟友在 9/11 事件前后大规模扩张的间谍机构，这再次引发了人们对大规模政府监控的长期担忧，而此前，前几代告密者、记者和公民自由主义者也曾提出过这种担忧。 美国国家安全局 (NSA) 和英国政府通信总部 (GCHQ) 赞助并受益于数据收集和分析学术研究和商业发展的急剧增长。 早期理解监控危险的方式要求对侵犯隐私的行为进行更严格的法律和技术分析，而越来越多的监控行为数据和复杂的分析技术成为可能。随着这些突出的干预措施的出现，对数据和算法的新中心地位的批判性诊断大量涌现。
In 2013, Safiya Noble, then a professor at the University of Illinois at Urbana-Champaign, now a MacArthur “Genius” fellowship awardee, published an excoriating look at Google search. “Commercial search implodes,” Noble wrote, “when it comes to providing reliable, credible, and historically contextualized information about women and people of color, especially Black women and girls.” Ostensibly bias-free technology quickly reproduced—and reinforced—racist and sexist biases toward Black women. “Continued study of these phenomena,” she wrote, “is an opportunity to contest the alleged neutrality of technology, while creating new opportunities for social justice and fair representation online.”5 In 2016, mathematician Cathy O’Neil described her own journey from academic to Wall Street employee to critic of data and algorithms gone unchecked. In her Weapons of Math Destruction, she explored how the incentives in data science undermined the humanity of its subjects: “The inclination is to replace people with data trails, turning them into more effective shoppers, voters, or workers to optimize some objective. This is easy to do, and to justify, when success comes back as an anonymous score and when the people affected remain every bit as abstract as the numbers dancing across the screen.”6 Without changing these incentives, data science, for all its promise, would dramatically alter the goals of organization after organization, from universities, to medicine, to social welfare, at the expense, most of all, of the least powerful members of society. Around the same time as these critical diagnoses, the revelations of the vast expansion of the spying apparatus of the United States and its “Five-Eyes” allies around 9/11 by Edward Snowden reanimated longstanding concerns about mass government surveillance, previously raised by earlier generations of whistleblowers, journalists, and civil libertarians. The American NSA (National Security Agency) and the British GCHQ (Government Communications Headquarters) sponsored and benefited from the dramatic explosion of academic research and commercial developments on the collection and analysis of data. Earlier ways of understanding the dangers of surveillance demanded a more critical legal and technological analysis of the violations of privacy that ever-greater amounts of data and sophistical analytical techniques made possible. A tremendous surge of critical diagnoses of the new centrality of data and algorithms emerged alongside these prominent interventions.
在批评声浪高涨的几年内，谷歌、Facebook 和 IBM 等公司都拥有了内部人工智能伦理学家。这些公司——在财务上相当于许多民族国家的经济强国——迅速吸收了批评运动，聘请了许多最杰出的批评家，但当批评声浪过大时，他们往往会噤声或剥夺他们的权力。这就像教皇聘请了马丁·路德，并把他安排在梵蒂冈的一个角落办公室里，而推动宗教改革的赦罪券则相对有增无减。谷歌等公司聘请的研究人员，如蒂姆尼特·格布鲁和玛格丽特·米切尔，经常发现自己面临着挑战，因为他们努力不被吸收。而计算机科学研究界则将这些对公平性的深切担忧转化为新的算法难题，但往往小心翼翼地将它们与权力的思考隔离开来。
Within a few years of the surge of critical concern, the likes of Google, Facebook and IBM all had in-house AI ethicists. These firms—economic powerhouses financially equivalent to many nation-states—had quickly appropriated the critical movement, hiring many of the most brilliant critics, but often silencing or disempowering them when the criticism got too hot. It was as though the Pope had hired Martin Luther and set him up in a corner office of the Vatican, while the indulgences spurring the Reformation went on relatively unabated. The researchers hired by the likes of Google, such as Timnit Gebru and Margaret Mitchell, too often found themselves challenged as they struggled not to be co-opted. The computer science research community, for its part, turned these deep concerns about fairness into new algorithmic puzzles, but too often carefully gated them away from reflections upon power.
瓦拉赫、诺布尔和奥尼尔等学者都敏锐地看到，新的算法系统如何以前所未有的速度和规模，在其自动化判断中轻易地复制了昔日的系统性不平等。新能力产生了新的权力。这些权力有可能使许多社会长期以来努力消除的许多不平等现象根深蒂固，尽管取得了参差不齐的成功。这些新技术往往强化了现有的结构性不平等和差异，普林斯顿大学教授鲁哈·本杰明将其描述为“一套产生社会关系模式的技术”，而这些模式本身“成为自然、不可避免、自动的黑箱”。7鲁哈认为，使用数据的技术所带来的客观性的外表有助于“编码”不平等。这种不平等延伸到构建算法系统的公司：正如前谷歌员工梅雷迪斯·惠特克所强调的那样，如今只有资源最丰富的公司和政府才真正具备大规模部署这些技术的能力。8
Scholars like Wallach, Noble, and O’Neil among many others all keenly saw how new algorithmic systems easily reproduced in their automated judgments the systemic inequalities of yore at an unprecedented rapidity and scale. New capabilities yielded new powers. These powers threatened to entrench many of the inequalities that so many societies had long struggled, with checkered and uneven success, to undo. Too often these new technologies reinforce existing forms of structural inequality and difference, in what Princeton professor Ruha Benjamin describes as “a set of technologies that generate patterns of social relations” that themselves “become black boxed as natural, inevitable, automatic.”7 The veneer of objectivity that comes from technologies using data serve, Ruha argues, to “encode” inequity. And this inequality extends into the firms building algorithmic systems: the capacity to deploy these technologies at vast scale today is only truly available to the best-resourced corporations and governments, as ex-Googler Meredith Whittaker has stressed.8
据说科幻小说作家威廉·吉布森曾说过：“未来已经到来——只是分布不太均匀。”我也相信这一点，但我认为他的意思恰恰相反。生活在低权利环境中的人们——贫困和工薪阶层社区、移民社区、有色人种社区、宗教或性少数群体——已经生活在数字化未来中，尤其是在高科技监控和纪律方面。
Science fiction writer William Gibson is believed to have said, “The future is already here—it’s just not very evenly distributed.” I believe that too, but in a way opposite to what I think he intended. People who live in low-rights environments—poor and working-class communities, migrant communities, communities of color, religious or sexual minorities—are already living in the digital future, especially when it comes to high-tech surveillance and discipline.
—弗吉尼亚·尤班克斯 9
—Virginia Eubanks 9
瓦拉赫、博伊德、奥尼尔、诺布尔以及其他学者和活动家的警告既不是第一个也不是唯一的，如今人们普遍认为，这些警告揭示了我们身处的数据洪流的弊端。正如丽莎·中村所说：“我在密歇根大学教授一门名为‘互联网是一场垃圾大火’的课程，我不需要向任何人解释这是什么意思。” 10
The warnings of Wallach, boyd, O’Neil, Noble, and other scholars and activists were neither the first nor the only examples of what is now widely recognized as illuminating the downsides of the data deluge in which we live. As Lisa Nakamura notes, “I teach a course at the University of Michigan called ‘the Internet is a trash fire,’ and I don’t have to explain to anybody what that means.”10
自 20 世纪 60 年代以来，早期学者和法律活动家就已指出数据积累和自动分析对隐私的危害，并指出此类调查往往会加剧现有的不平等现象。社会学家对数据赋能算法对民主政治的影响表示担忧。Zeynep Tufekci 在 2014 年警告说，“那些拥有资源和渠道的人利用这些工具在民主国家开展高度有效、不透明和不负责任的说服和社会工程活动的能力正在增强”。政治、公民和商业领域。” 11虽然政治说服和量化或“绩效”营销在 2014 年并不是什么新鲜事，但它们与政治影响力运作和“微目标定位”（即优化向个人传递不同数字信息的能力）相结合，开辟了将选民的现实状况分割成蕾妮·迪雷斯塔 (Renée DiResta) 后来所说的“定制现实”的可能性。12 2016 年美国总统大选让人们意识到这种微目标的现实，尤其是剑桥分析公司及其利用 Facebook 收集的数据。Facebook 上看似无害的性格测试被用来“操纵公众”，正如 Tufekci 所说，这个故事很好地反映了公众的想象和恐惧，导致 Facebook 首席执行官马克·扎克伯格于 2018 年春季在国会作证，向议员们保证，尽管选民们担心，但他们的数据不会被出售，门洛帕克一切安好。（然而，议员们的回应并不总是让选民感到放心：一位困惑的参议员问 Facebook 是否“与”Twitter“相同”，另一位参议员要求扎克伯格澄清该公司是否销售广告——这是他们的主要业务。）
Early generations of scholars and legal activists from the 1960s onward had signaled the dangers of the accumulation and automatic analysis of data to privacy and noted how such inquiry often exacerbates existing inequalities. Sociologists had raised concerns about the impact of data-empowered algorithms on democratic politics. Zeynep Tufekci warned in 2014 of the rising “capacity of those with resources and access to use these tools to carry out highly effective, opaque and unaccountable campaigns of persuasion and social engineering in political, civic and commercial spheres.”11 While political persuasion and quantitative or “performance” marketing were not new in 2014, their combination with political influence operations and “microtargeting”—the ability to optimize the delivery of different digital messages to individuals—opened up the possibility of fracturing the realities of the electorate into what Renée DiResta later called “bespoke realities.”12 The US presidential election of 2016 brought home the realities of such microtargeting, especially around the company Cambridge Analytica and its exploitation of data gleaned via Face-book. Public imagination and fears were well captured by the story of what appeared to be a harmless personality quiz on Facebook being leveraged to “engineer the public,” as Tufekci put it, leading to congressional testimony by Facebook’s CEO Mark Zuckerberg in the spring of 2018, reassuring lawmakers that, despite the fears of their constituents, their data was not for sale and all was well in Menlo Park. (The lawmakers’ responses were not always reassuring to the electorate, however: one confused senator asked if Facebook was “the same as” Twitter, and another senator asked Zuckerberg to clarify if the firm sells ads—their primary business.)
人们担心算法会在整个民族国家范围内推动“计算政治”，这与人们对算法会创造弗吉尼亚·尤班克斯所说的“数字救济院”的担忧如出一辙。尤班克斯 2018 年出版的《自动化不平等》一书追溯了三个关于算法剥夺穷人和有需要的人权力的故事。她分享了国家使用的不透明算法（有的复杂，有的简单）如何加剧社会不平等，给那些最无权自卫和批评算法使用的人造成大规模伤害。她的一位线人警告我们，除了最有权力的人之外，不公正的伤害可能会用在所有人身上，她责备道：“你应该注意发生在我们身上的事情。你是下一个。” 13 个预测性警务的故事在美国及其他地区的审判中，反乌托邦科幻小说变成了现实，其中最臭名昭著的或许是芝加哥的“战略对象名单”。14当警察部署数据赋能算法时，它们已经导致了错误逮捕和监禁，人们越来越担心算法设计和部署中的偏见。卫生工作者和公共卫生官员都面临着骚扰、否认和反叙事。民族国家面临着大规模的不稳定和虚假信息。
Fears of the power of algorithms to drive “computational politics” at the scale of entire nation-states mirrored concerns raised about the power of algorithms to create what Virginia Eubanks calls a “digital poorhouse.” Her 2018 book Automating Inequality traces three stories about algorithms disempowering the poor and the needy. She shares how opaque algorithms used by the state—some complex, some simple—serve to exacerbate social inequality, delivering harms at scale to those least empowered to defend themselves and critique their use. Warning us of the way unjust harms can be used on all but the most empowered, one of her informants chides her: “You should pay attention to what happens to us. You’re next.”13 Tales of predictive policing have moved from dystopian science fiction to reality in trials in the United States and beyond, perhaps most notoriously Chicago’s “Strategic Subjects List.”14 When data-empowered algorithms are deployed by police forces, they have already led to wrongful arrests and incarceration, with growing concern about biases in their design and deployment. Health workers and public health officials alike face harassment, denial, and counternarratives. Nation-states face destabilization and disinformation at enormous scale.
算法决策系统充斥着越来越多人的详细数据，其带来的潜在威胁在许多方面都是新的。它们使政府和企业能够以全新的规模了解我们的日常活动：以前针对小群体（通常是最边缘化或持不同政见者）的技术可以应用于整个人口。它们构成了一种前所未有的亲密关系，因为它们为我们的人际交流、新闻和信息来源提供动力，甚至通过算法调节我们的关系。15这使得此类系统（包括推荐电影和娱乐、新闻或恋爱对象的算法）在滥用或设计不当的情况下更具破坏性。例如，在虚假信息方面，开放信息平台的性质意味着数据赋能算法的危险不仅来自民族国家，也来自我们的邻国。
The potential threats from algorithmic decision systems replete with granular data on increasingly large numbers of people are in many ways new. They enable governments and corporations to know about our everyday activity at an entirely new scale: techniques previously directed at small groups, often the most marginalized or dissident, can be applied to the entire population. They constitute an unprecedented intimacy in that they power our interpersonal communications, our sources of news and information, and even algorithmically moderate our relationships.15 This makes such systems (including algorithms recommending movies and entertainment, news, or romantic partners) all the more potentially damaging in the cases of either abuse or poor design. In the case of disinformation, for example, the nature of open information platforms means that the dangers of data-empowered algorithms come from not only nationstates but our neighbors.
在学术界，人们对机器学习系统爆炸式增长的反应各不相同，既有热情也有恐慌，各种技术专家、社会科学家和人文学者的参与度也越来越高。然而，产业与学术界之间日益紧密的关系理所当然地引发了人们的愤怒：随着赞助的工业研究规模不断扩大，与传统的政府资助相媲美，一种“俘获”得以实现，其中批评科技公司的研究受到阻碍或仅仅因为害怕失去经济支持而失去动力。16
Within academia, responses to the explosion of machine learning systems are widely varied, between enthusiasm and alarm, with growing participation from varied technologists, social scientists, and humanists. Yet the tightening relationship between industry and academia rightly raises hackles: as the size of sponsored industrial research grows, rivaling that of traditional government funding, a type of “capture” is enabled in which research critical of technology companies is thwarted actively or merely disincentivized by the fear of lost financial support.16
在这个产学研联合体之外，在活动家以及没有直接或从数据赋能的科技公司获得经济利益的院系和教职员工中，不仅存在担忧，而且公开采取行动限制科技公司的权力，要么通过倡导国家监管，要么以“私人订购”的形式，包括劝阻他人为危害社会的科技公司工作或与科技公司合作。17特别是在过去几年中，企业以各种方式作出了回应，有的采用旧的，有的采用新的。传统的回应，如政府游说和面向公众的公关活动，在强度和财务规模上都有所增加，甚至变得狡猾。针对算法担忧的具体回应包括以“公平”为名的技术修复以及建立“人工智能伦理”团队、角色或负责人。这两种做法都引起了美国国会、关注者和批评者的不同反应。这两种做法都未能成功地彻底改变最强大公司的内部流程。法学教授弗兰克·帕斯夸尔多年来一直警告说，大公司会利用透明度等真正重要的价值观。 2016 年，一个研究者联盟提出了重要的“可信赖算法原则”，其中指出，
Outside this industrial-academic complex, among activists and among departments and faculty not directly or financially benefiting from data-empowered technology companies, there is not just concern but open activism to limit the power of technology companies, either by advocating for state regulation or in the form of “private ordering,” including discouraging others from working for or with technology companies that harm society.17 Corporations have responded, particularly over the past few years, in a variety of ways, some old and some new. Traditional responses, such as government lobbying and public relations campaigns to the public, have grown in intensity and financial scale—and even deviousness. Responses particular to concerns over algorithms include both technical fixes under the label “fairness” as well as establishing “AI ethics” teams, roles, or principals. Both have drawn mixed responses from the US Congress, from the concerned, and from the critical. Neither has yet succeeded in dramatically changing internal processes at the most powerful of companies. Law professor Frank Pasquale has warned for many years that enormous corporations co-opt genuinely important values such as transparency. In 2016, a coalition of researchers produced important “Principles for Accountable Algorithms” that argued,
自动决策算法现已被整个行业和政府广泛使用，支撑着从动态定价到就业实践再到刑事判决等许多流程。……在此背景下的问责包括报告、解释或证明算法决策的义务，以及减轻任何负面的社会影响或潜在危害。18
Automated decision making algorithms are now used throughout industry and government, underpinning many processes from dynamic pricing to employment practices to criminal sentencing. . . . Accountability in this context includes an obligation to report, explain, or justify algorithmic decision-making as well as mitigate any negative social impacts or potential harms.18
我们需要做的还远远不止这些，我们现在所提到的大多数作者可能都会同意这一点；我们必须拥有强大的制度形式，能够追究责任，而不仅仅是强制出具报告。应对算法系统的危险和前景需要集中的政治行动，能够影响数据赋予谁权力，不赋予谁权力；它需要清楚地了解我们目前的状况是多么偶然——而不是一成不变。我们越了解这些系统的起源，我们就越有能力共同抗争、反抗，并将它们用于更公正的用途。19
Far more is necessary, as most of the writers we invoke would probably now agree; we must have robust institutional forms that enable a holding to account, not simply force the production of an account. Tackling the dangers and promise of algorithmic systems demands concentrated political action, capable of affecting who data empowers and who it does not; and it demands a clear understanding of how contingent—how not set in stone—our current state of affairs is. The better we understand the genesis of those systems, the better equipped we collectively will be to contest, defy, and put them to more just uses.19
历史与批评
History and Critique
我们以一条技术定律作为本章的开篇，即克兰兹伯格第一定律，它以一位技术史学家的名字命名，他在 1986 年写道：
We opened the chapter with a law of technology, Kranzberg’s first law, named after a historian of technology who wrote in 1986:
我的第一定律——技术既不是好的也不是坏的；它也不是中性的——应该不断提醒我们，历史学家的职责是比较短期结果与长期结果、乌托邦的希望与现实、可能发生的事情与实际发生的事情，以及各种“好”与可能的“坏”之间的权衡。所有这些都只能通过观察技术如何以不同的方式与不同的价值观和制度，甚至与整个社会文化环境互动来实现。20
my first law—Technology is neither good nor bad; nor is it neutral—should constantly remind us that it is the historian’s duty to compare short-term versus long-term results, the utopian hopes versus the spotted actuality, the what-might-have-been against what actually happened, and the trade-offs among various “goods” and possible “bads.” All of this can be done only by seeing how technology interacts in different ways with different values and institutions, indeed, with the entire sociocultural milieu.20
如今，对数据的担忧日益增加，乐观的声音也随之而来，而乐观的声音则被利用数据理解世界和促进日常活动的技术进步所激发。在我们写作时，我们接受的语音转文本和自动拼写检查水平远远超过几年前的最佳努力也已不复存在。在研究方面，我们已经看到预测蛋白质折叠或从临床图像数据和基因组数据中识别疾病的能力取得了显著进步。自动驾驶汽车和个性化或“精准”医疗的美好未来承诺充斥着科技媒体和科技公司的营销材料。我们不需要接受围绕数据的过度营销炒作，就能认识到这些嵌入我们社会、政治和经济体系的技术所带来的深远影响，其中许多影响是无意的。
The growing voices of alarm around data today compete with voices of optimism, fired by clear technological advances in using data to make sense of our world and to facilitate everyday activities. As we write, we accept as normal a level of speech-to-text and automated spell-checking that far exceeds the best efforts of even a couple of years ago. In research, we’ve seen marked advances in the ability to predict protein folding, or to identify disease from clinical image data and genomic data. Promises of great futures with self-driving cars and personalized or “precision” medicine fill the tech press and the marketing materials of tech companies alike. And we need not accept the overblown marketing hype around data to recognize the profound effects, many unintentional, wrought by these technologies embedded within our social, political, and economic systems.
这些发展给许多权威机构和许多职业带来了挑战，从科学家到广告商，从医生到律师。机器大规模取代手工工人的故事与工业革命一样古老；现在机器正在取代精英白领——而且速度很快。例如，医生越来越意识到机器将很快完成更多重要的诊断实践：“预测颠覆最有可能发生在何处很难，但如果今天感觉很常规，那么明天它很可能成为机器的目标。放射学等临床专业可能不会消失，但它们肯定会发生重大转变。” 21即使这些传统职业面临挑战，也需要全球范围内的新工人来使系统能够完成工作。
With these developments come challenges to many authorities and many professions, from scientists to advertisers, physicians to lawyers. The story of machines replacing hand workers at a vast scale is as old as the Industrial Revolution; and now the machines are coming for elite white-collar workers—and quickly. Doctors, for example, are increasingly recognizing that machines will soon do more central diagnostic practices: “Predicting where disruption is most likely to occur is hard, but if it feels routine today, then it is likely to be a target for the machine tomorrow. Clinical specialties like radiology might not disappear, but they certainly will be heavily transformed.”21 Even as these traditional professions are challenged, a global range of new workers will be required to enable the systems doing the work.
在过去十年中，我们看到企业和政府对个人权利、伤害和正义的威胁日益增加；与此同时，我们也目睹了个人生活和研究的巨大发展，以及未来技术优势的前景。现在也很清楚，那些掌权者——尤其是国家和企业掌权者——不会在没有强大压力和倡导的情况下放弃数据赋能的能力。我们必须接受 Kranzberg 的挑战，了解短期和长期结果，并了解我们的选择构成了权力的大大小小重新排序。
In the last ten years, we’ve seen increasing threats to individuals’ rights, harms, and justice from corporations and governments alike; at the same time, we’ve also witnessed great developments, in our personal lives, and in research, and the promise of technological benefits to come. It is also, by now, clear that those in power—particularly state and corporate power—will not be giving up on data-empowered capabilities without intense pressure and advocacy. We must take on Kranzberg’s challenge to understand the short- and long-term results, as well as to understand the ways our choices constitute small and large reorderings of power.
历史作为溶剂
History as Solvent
强大的权力往往不愿意探究使它们成为可能甚至占据主导地位的历史起源。复杂的历史扰乱了它们权力的显而易见性和合法性。22在审视技术崛起的那些不那么显而易见的方式时，历史扰乱了某些技术本身的发展推动历史的观点，这种观点被称为“技术决定论”。例如，对于许多感兴趣的参与者来说，声称旧的隐私观在互联网时代已经过时，甚至互联网本身导致了隐私的衰落，是非常有利可图的。这两种说法都不正确。但这样的说法提供了一个强有力的历史版本，在围绕互联网的辩论中无处不在，它使当前的秩序合法化。
Powerful forces often are reticent to investigate the historical genesis that made them possible—or even dominant. Complex histories unsettle the obviousness, the legitimacy, of their power.22 In looking at the far from obvious ways technologies come to prominence, history unsettles the idea that the growth of certain technologies themselves drives history, a view called “technological determinism.” It has been very lucrative for many interested actors, for example, to claim that older views of privacy are outdated in the age of the internet, even that the internet itself causes the decline of privacy. Neither claim is true. But such stories offer a potent version of history, ubiquitous in debates around the internet, that legitimates the current order of things as necessarily so.
历史可以沦为对更人道、更美好过去的怀念，但它不必如此。无论当代算法决策的新颖性、危险性和规模如何，使用量化指标的无灵魂官僚机构的出现往往有着一段黯淡的历史。从米歇尔·福柯 (Michel Foucault) 和伯纳德·科恩 (Bernard Cohn) 到杰奎琳·韦尼蒙特 (Jacqueline Wernimont)、玛莎·霍德斯 (Martha Hodes)、西蒙娜·布朗 (Simone Browne) 和哈利勒·纪伯伦·穆罕默德 (Khalil Gibran Muhammad) 等学者表明，从 19 世纪初开始，对人的量化就有着悠久的历史，包括对学生、种族、殖民地人民、奴隶、士兵、穷人、精神病患者和囚犯进行排名和分类。23萨拉·伊戈 (Sarah Igo)、伊曼纽尔·迪迪埃 (Emmanuel Didier)、丹·布克 (Dan Bouk) 和艾米丽·麦钱特 (Emily Merchant) 等历史学家探索了调查和人口普查不仅仅是记录：它们构成了公众和人口；它们促成了团结的形式和政府行动的类型——以及不作为。数据是通过数据产生的，而获取和分析这些信息的过程往往会戏剧性地循环往复，影响受到官方审查的人。24
History can collapse into nostalgia about a more humane, better past, but it need not. Whatever the novelty, dangers, and scale of contemporary algorithmic decisionmaking, the emergence of soulless bureaucracies using quantitative measures has an often-dim history. Scholars from Michel Foucault and Bernard Cohn to Jacqueline Wernimont, Martha Hodes, Simone Browne, and Khalil Gibran Muhammad show how the quantification of peoples has a long history from the early nineteenth century onward in ways of ranking and classifying students, races, colonized peoples, enslaved people, soldiers, the poor, the mentally ill, and the incarcerated.23 Historians like Sarah Igo, Emmanuel Didier, Dan Bouk, and Emily Merchant explore how surveys and censuses don’t simply record: they constitute publics and populations; they enable forms of solidarity and types of governmental action—and inaction. Data is made, not found, and the process of procuring and analyzing it often dramatically loops back to shape the people under official scrutiny.24
早在 SAT 将大学申请者分数化之前，心理学家查尔斯·斯皮尔曼 (Charles Spearman) 就提出了一个数学上的“一般智力”分数，将智力简化为一个数字；早在作家和亚马逊 (Amazon) 能够知道有多少在线读者在阅读他们的作品之前，19 世纪的机械工程师弗雷德里克·泰勒 (Frederick Taylor) 就引入了科学管理来量化工人的产出；历史学家凯特琳·罗森塔尔 (Caitlin Rosenthal) 指出，在这两者出现的几十年前，复杂的会计和簿记手段已成为种植园奴隶制的核心。25然而，严谨的定量研究可以为组织我们的社会、医疗和政治生活带来帮助，使我们的社会受益匪浅。举个例子，我们对疫苗的信任关键在于评估功效和衡量危害的标准化定量过程。量化指标可以（而且已经）提供问责制，但它们却以惊人的速度对我们不利。
Long before the SAT reduced college applicants to a score, the psychologist Charles Spearman proposed a mathematical “general intelligence” score to reduce intellect to a number; long before writers and Amazon could know how many online readers were engaging with their writing, the nineteenth-century mechanical engineer Frederick Taylor introduced scientific management to quantify worker output; many decades before both, sophisticated means of accounting and bookkeeping were at the center of plantation slavery, as historian Caitlin Rosenthal has shown.25 And yet our societies benefit from the knowledge that rigorous quantitative studies can bring in organizing our social, medical, and political lives. Our trust in vaccines—to take one example—rests crucially upon a standardized quantitative process of assessing efficacy and gauging harm. Quantified measures can— and have—and do—provide accountability, but they’ve been turned against us with a feverish pace.
数值问责制的主要方法在很大程度上是作为抵制专家判断的工具而出现的，这些判断隐藏在政府、教育和企业实体中。科学史学家西奥多·波特认为，标准化的数值会计形式（如成本效益分析）的出现是为了挑战人类“黑匣子”的权力，这些专家的权威建立在传统地位和不透明的判断形式之上。数值问责制承诺透明度和受规则约束的客观性，通常在对专家不信任的情况下获得重视。26例如， 1933年的“证券真实性”法案试图通过统一的会计和报告标准来加强对资本市场的信任，但遭到了华尔街及其会计师的强烈抵制。迫使银行在多年后披露贷款协议的行为暴露了系统性的种族主义标准和强制制定的新标准。无论其局限性如何，自 19 世纪以来，使用量化指标提供有关主要机构的信息一直是制约国家和企业权力的强大工具，尤其是制约组织中专家决策的不透明性。这些知识可以让人们挑战专家，确保公平，使决策透明。
Key methods of numerical accountability emerged in large part as tools for resisting expert judgment ensconced in critical governmental, educational, and corporate entities. The science historian Theodore Porter argues that standardized forms of numerical accounting, such as cost-benefit analysis, arose to contest the power of human “black boxes,” experts with authority grounded in traditional status and opaque forms of judgment. Numerical accountability, with its promises of transparency and rule-bound objectivity, typically gained prominence in conditions of distrust about experts.26 The 1933 “Truth in Securities” act, for example, sought to bolster trust in capital markets through uniform standards of accounting and reporting, which were bitterly resisted by Wall Street and its accountants. Forcing banks years later to reveal their protocols for lending money revealed systematically racist criteria and forced creation of new criteria. Whatever its limits, making information about major institutions available using quantitative measures has been, since the nineteenth century, a formidable tool to check state and corporate power, particularly the opacity of expert decisionmaking in organizations. Such knowledge can enable people to challenge experts, to ensure fairness, to make decisions transparent.
然而，让强大机构更容易被公众接受的机制长期以来一直专注于公众。算法系统并没有让强大的机构对我们透明，而是越来越多地将我们暴露在强大的机构面前。27在过去的四十年里，企业、大学和政府以越来越快的速度对个人员工和公民实施了数字问责制。实施这样的衡量体系并非不可避免或微不足道。28在过去的四十年里，人们见证了在具体情况下采用和应用数字问责制的技术激增。经过数十年的努力，数字测量现在渗透到从工厂到大学、从优步司机到渔场的各个工作场所，通过指标体系，雇主可以了解员工，并对员工的几乎所有活动向老板负责。算法——通常是秘密的、专有的、很少透明的——处理这些数据以进行排名和分类、晋升和解雇、奖励和惩罚。问责制度往往强加于制度体系的较低层级，可预测地沿着社会经济和种族界限划分。这些措施并没有让强大的机构对外透明，反而往往让日常员工和公民对强大的机构透明。它们确实做到了这一点。通过通常不受审查的算法决策和分类。
And yet the mechanisms so useful for making powerful institutions more accessible to the public have long since been focused on the public. Rather than making powerful institutions transparent to us, algorithmic systems increasingly lay us bare to powerful institutions.27 In the last four decades, corporations, universities, and governments have imposed numerical measures of accountability on individual employees and citizens at an ever-accelerating pace. Nothing is inevitable or trivial about imposing such systems of measure.28 The last forty years have seen an explosion of techniques for insisting on numerical accountability adapted and applied in concrete contexts. Thanks to decades of effort, numerical measurement now saturates workplaces, from factories to universities, from Uber drivers to fisheries, through systems of metrics that make employees knowable to employers—and accountable to their bosses for nearly all their activities. Algorithms—often secret, proprietary, and rarely transparent—process this data to rank and to classify, to promote and to fire, to reward and to punish. Systems of accountability tend to impose themselves on lower rungs of institutional systems, falling predictably along socioeconomic and racial lines. Rather than rendering powerful institutions transparent to outsiders, these measures instead often render everyday employees and citizens transparent to powerful institutions. And they do so through algorithmic decision-making and classification usually removed from scrutiny.
我们将从谷歌搜索和 Uber 拼车出现之前，讲述比利时天文学家的梦想。他发明了体重指数和“社会物理学”。我们将看到，他的目标听起来很现代：利用当今最新的技术收集个人数据。并赋予那些使用这些技术的人权力，以改善社会本身。
We’ll start, long before Google search and Uber rideshares, with the dreams of the Belgian astronomer who invented the body mass index and “social physics.” His goals, we’ll see, sound contemporary: to use the latest technologies of the day and collected data about individuals. And to empower those using these technologies to improve society itself.
Social Physics and l’homme moyen
19 世纪的比利时天文学家阿道夫·凯特勒启发了弗洛伦斯·南丁格尔。这位先知对制定法律有着新的看法，具有基于证据的眼光。这位比利时人建议，写下“你对某项立法的期望。x 年后看看它是否满足了你的期望，哪些地方没有达到你的期望。”南丁格尔在 1891 年写道，她抱怨说，她那个时代的立法不涉及这些数据：“你们如此迅速地改变法律和执行法律，没有调查过去和现在的结果，以至于一切都只是实验、跷跷板、教条主义，就像两个羽毛球运动员之间的羽毛球。” 1南丁格尔解释说，这位比利时人是“全世界最重要科学的创始人”，他提供了“所有政治和社会管理所必需的一门科学”。凯特勒“在世时没有看到它以任何实际方式明显影响政治家风范（没有它就没有政治家风范）或政府。” 2虽然这门新科学——这些新能力——还没有重新排列权力，但南丁格尔确信它们应该会重新排列。我们的世界就涉及这样的权力重新排列。
A nineteenth-century Belgian astronomer, Adolphe Quetelet, inspired Florence Nightingale. A prophet with a new vision for making law. An evidence-based vision. Write down, this Belgian advised, “what you expect from such and such legislation. After x years see where it has given you what you expected and where it has failed.” Writing in 1891, Nightingale complained that lawmaking of her times involved no such data: “You change your laws and your administering of them so fast and without inquiry after results past and present, that it is all experiment, see-saw, doctrinaire, a shuttlecock between two battledores.”1 This Belgian, the “founder of the most important science in the whole world,” Nightingale explained, had provided “the one science essential to all political and social administration.” Quetelet “did not live to see it perceptibly influence, in any practical manner, statesmanship—of which there is none without it—or government.”2 While this new science—these new capabilities—had not yet rearranged power, Nightingale was sure that they should. Our world involves just such a rearrangement of power.
凯特勒给了我们体重指数，即统计学上的平均人格概念，最重要的是，他极大地改变了我们对社会的看法。哲学家伊恩·哈金打趣道，凯特勒“喜欢数字，而且乐于得出结论。” 3
Quetelet gave us the body mass index, the idea of the statistically average person, and, above all, dramatically altered how we think about societies. Quetelet “was fond of numbers,” the philosopher Ian Hacking quipped, “and happy to jump to conclusions.”3
尽管凯特勒对社会有着许多激进的新想法，但他还是想避免激进的破坏。从法国和海地革命到拿破仑帝国，他那个时代已经经历了太多这样的破坏。1830 年，革命者占领了他在布鲁塞尔的新天文台，令他非常沮丧。凯特勒在给一位朋友的信中写道：“我们的天文台刚刚被改造成一座堡垒。” 4暴力革命已经失去了吸引力。凯特勒痴迷于将数学应用于社会，他试图创造一门非革命性变革的新科学。他解释说，在政治和社会生活中，“突然行动”会导致权力的浪费。“这一原则对革命党人来说是有利的，”他指出。社会需要改革。革命不是出路。“突然行动总是会造成一定程度的生命力损失。这一原则对革命党人来说是不利的，除非他们将权力推向更有用的方向[并且]同意失去一部分[这些权力]。” 5
For all his radical new ideas for thinking about society, Quetelet wanted to avoid radical disruption. His time had seen far too much of that, from the French and Haitian revolutions to the Napoleonic Empire. In 1830, revolutionaries occupied his new astronomical observatory in Brussels, to his great dismay. “Our observatory,” Quetelet wrote a friend, “has just been converted into a fortress.”4 Violent revolution had lost its appeal. Obsessed with applying mathematics to society, Quetelet sought to create a new science of nonrevolutionary change. In political as well as social life, he explained, “abrupt movement” causes a wasteful loss of force. “This principle is advantageous to the partisans of a revolution,” he noted. Society needed reform. Revolution wasn’t the way. “Abrupt movements are never made without a certain loss of live force. This principle is not advantageous to the partisans of a revolution, unless they impel forces in a more useful direction [and] consent to lose a portion [of these forces].”5
一种基于有关人的数据的新的科学政治应该重新安排权力。逐步地。不占用建筑物。不中断。
A new scientific politics, based on data about people, should rearrange power. Gradually. Without occupying buildings. Without disruption.
官僚主义、预算问题和施工挑战长期推迟了天文台的完工。在等待观测天空的同时，凯特勒利用研究星星的最佳技术来思考对人的观察。6凯特勒直接受到十八世纪末天地物理和天文模型的成功影响，并对十九世纪初欧洲政治和军事权力的动荡感到沮丧，因此试图创造一种新的“社会物理学”。但在 1830 年，数字并不是理解人性或权力关系的明显方式——南丁格尔几十年后呼吁社会物理学并非偶然。
Bureaucracy, budget problems, and construction challenges long delayed the completion of his observatory. While he waited to survey the sky, Quetelet drew on the best techniques for studying the stars to think through observations about people.6 Directly influenced by the triumphant late eighteenth-century successes of physical and astronomical models of heaven and earth, and dismayed by the early nineteenth-century political and martial upheavals of power in Europe, Quetelet sought to create a new “social physics.” But numbers were not the obvious way to understand humanity or power relations in 1830—it was no accident that Nightingale was calling for social physics many decades later.
粗俗的统计数据
Vulgar Statistics
我们怎么会认为数字对于理解世界及其人民的生活至关重要呢？从艺术家到人类学家，从小说家到大维齐尔，批评家们长期以来一直拒绝量化。一位德国辩论家在 1806 年写道：“这些愚蠢的家伙传播了一种疯狂的想法，即只要知道一个国家的规模、人口、国民收入和周围吃草的愚蠢野兽的数量，就可以了解这个国家的实力。” 7他认为，真正的统计数据，对国家的真正了解，不像它的“粗俗”表亲，需要仔细描述和了解历史。这种调查超越了物质，以掌握不同国家的道德和精神结构。自 17 世纪末以来，为道德指导、新闻和利润而统计死亡率的做法有所增加，但毫无技巧地将这种粗糙的表格应用于国家治理的重大问题是一种诅咒。数字处理者是“表格统计学家”，而不是真正的统计学家。数字描述“不涉及国家、道德、神灵的精神权力和关系”。这些统计学家“根本看不到质量，只看到数量。” 8
How did we come to think that numbers are essential to understand the world and the lives of its peoples? From artists to anthropologists, from novelists to grand viziers, critics have long said no to quantification. “These stupid fellows,” a German polemicist wrote in 1806, “disseminate the insane idea that one can understand the power of a state if one just knows its size, its population, its national income, and the number of dumb beasts grazing around.”7 Real statistics, genuine knowledge of the state, he maintained, unlike its “vulgar” cousin, involved careful description and knowledge of history. Such investigation transcended the material to grasp the moral and spiritual texture of different countries. Tabulating mortality for moral guidance, news, and profit had grown since the late seventeenth century, but artlessly applying such crude tables to major questions of statecraft was anathema. Number crunchers were “table-statisticians,” not real statisticians. Numerical depiction “does not touch upon the spiritual forces and relationships of states, morals, the divine.” Such statisticians “see quality not at all, but only quantity.”8
两百年后，前共和党演讲撰稿人、《华尔街日报》专栏作家佩吉·努南（Peggy Noonan）也同样谴责了她认为最荒谬的状况：
Two hundred years later, former Republican speech-writer and Wall Street Journal op-ed contributor Peggy Noonan similarly decried a state of affairs she found most ridiculous:
前几天，一位共和党政坛老手转发给我一份奥巴马 2012 年竞选团队的招聘启事。读起来就像是火星人搞的政治。“分析部门”正在寻找“预测模型/数据挖掘”专家加入竞选活动的“多学科统计学家团队”，该团队将使用“预测模型”来预测选民的行为。9
The other day a Republican political veteran forwarded me a hiring notice from the Obama 2012 campaign. It read like politics as done by Martians. The “Analytics Department” is looking for “predictive Modeling/Data Mining” specialists to join the campaign’s “multi-disciplinary team of statisticians,” which will use “predictive modeling” to anticipate the behavior of the electorate.9
到 2016 年，双方在这方面都拥有强大的数据运营能力。
By 2016 both parties had formidable data operations in this vein.
数字并不总是理解和行使权力的明显方式。它是如何变成这样的？为什么我们现在求助于它们？一旦被计算机化，它们是如何成为病态的，同时又是解放的？关于人和事物的数据的数学分析是如何成为理解和控制世界、预测和规定的主导方式的？启蒙运动末期对数值统计的批评者很清楚，数据是极其人造的。正如丽莎·吉特尔曼几年前指出的那样，“原始数据是一个矛盾的说法”，因为所有数据收集都来自人类对收集什么、如何分类、包括谁和排除谁的选择；所有收集都涉及认知偏见和对这些信息进行分类、存储和处理截然不同的基础设施。10数据是制造出来的，而不是被发现的，无论是在 1600 年、1780 年还是 2022 年。11这些数据是如何变得强大的？收集、存储和分析这些数据的结构是如何建立的？使用它的论点为何变得如此令人信服——甚至具有法律必要性？
Numbers haven’t always been the obvious way to understand and to exercise power. How did it get that way? Why do we now turn to them? And once computerized, how are they pathological, as well as liberating? How did the mathematical analysis of data about people and things come to be such a dominant way to understand and to control the world, to predict and to prescribe? The critics of numerical statistics at the end of the Enlightenment well understood that data is profoundly artificial. As Lisa Gitelman noted some years ago, “raw data is an oxymoron,” as all data collection comes through human choice about what to collect, how to classify, who to include and to exclude; all collection involves cognitive biases and radically different infrastructures for categorizing, storing, and processing that information.10 Data is made not found, whether in 1600 or 1780 or 2022.11 How did such data become powerful? How did the structures to collect, to store, and to analyze it get built? How did arguments using it become so convincing—and even legally necessary?
在 18 世纪的欧洲，战争、税收，有时甚至是生命，通常是死亡，占据了统治者关注的焦点。18 世纪的欧洲，流血事件不断，中间夹杂着和平，经常蔓延到美洲和其他地区的残酷冲突。战争需要金钱；金钱需要税收；税收需要不断壮大的官僚机构；而这些官僚机构需要数据。启蒙运动欧洲的新兴国家需要知道需要哪些资源他们拥有：人口、土地、贵金属和行业。统计最初是关于国家及其资源的知识，没有任何特别的量化倾向或对洞察力（预测或其他）的渴望。从 1780 年开始，计数开始爆发，伊恩·哈金 (Ian Hacking) 将其描述为“数字的雪崩” 。12
In Europe in the eighteenth century, war, taxes, and sometimes life and usually death dominated the concerns of rulers. Eighteenth-century Europe saw continual bloodshed with punctuations of peace, often extending to brutal conflict in the Americas and elsewhere. War required money; money required taxes; taxes required growing bureaucracies; and these bureaucracies needed data. The burgeoning states of Enlightenment Europe needed to know what resources they had: people, land, precious metals, and industries. Statistics was originally knowledge of the state and its resources, without any particularly quantitative bent or aspirations at insights, predictive or otherwise. From 1780, an explosion of counting took off, which Ian Hacking memorably described as “an avalanche of numbers.”12
这种新的、高度数字化的统计数据威胁到了理解统治和理解人民的旧方式。新统计数据的倡导者没有以政治哲学经典为基础来组织国家，也没有以古代和现代国家的历史为指导，而是专注于与指导统治者相关的“土地和人民”的量化描述。改革派官员掌握了研究人民和国家的新方法，试图说服统治者，他们及其方法对于国家的发展和健康必不可少。他们试图描述和解释这些描述，以提供政策建议。人口普查从来都不是“中立的”，而是以目标为设计目标，并以建议政策的方式进行解释，特别是资源分配。十八世纪末，新成立的美利坚合众国将人口普查写入其最基本的法律《宪法》。当时，和现在一样，数字是政治性的。
This new, highly numerical statistics threatened the older ways of understanding rule and understanding people. Rather than basing the organization of the state upon the classics of political philosophy and using the history of states ancient and modern as guides, advocates of the new statistics focused on the quantitative descriptions of “land and people” relevant to guiding the ruler. Reforming officials, armed with new ways of studying the people and the state, tried to convince rulers that they and their methods were necessary for the growth and health of the state. They sought to describe, and to interpret, these descriptions as providing suggestions for policy. Enumeration was never “neutral” but designed with goals in mind, and interpreted in ways that suggested policy, particularly allocation of resources. Near the close of the eighteenth century, the new United States of America enshrined the census in its most fundamental law, the Constitution. Then, as now, numbers were political.
个人数据的历史（即其收集和解释）往往涉及政治、军事、殖民和工业权力的强大强化。鉴于收集中国、印加空间和其他地方的土地和人民信息的悠久传统，这种做法并非启蒙运动后期欧洲国家所独有。然而，从 18 世纪到 20 世纪，量化在欧洲、然后是美国和世界各地的殖民地获得了全新的中心地位。13
The history of personal data—its collection and interpretation—often involves the powerful reinforcing of political, military, colonial, and industrial power. Given long traditions of collecting information about lands and peoples in China, in Incan space and elsewhere, such practices were not exclusive to late Enlightenment European states. Yet quantification gained a radical new centrality in Europe and then the United States and colonies worldwide from the eighteenth to the twentieth century.13
统计最初是国家在工业、商业和军事不断发展的时期采用的一项新技术。竞争。马尔萨斯的继承者们，我们担心人口过剩。相反，十八世纪的欧洲思想家们担心人口不足，常常将经济欠发达归咎于此。君主及其顾问们开始将国家和“种族”的实力视为由人口规模和活力来量化。
Statistics was initially a new technology for states at a moment of increasing industrial, commercial, and martial competition. Heirs of Malthus, we worry about overpopulation. European thinkers of the eighteenth century, on the contrary, were anxious about underpopulation, often attributing economic underdevelopment to it. Monarchs and their advisors came to view the strength of states—and of “races”—as quantified by the size and vigor of its population.
定期发布的记录教区死亡原因的法案构成了 17 世纪英国最早可识别的数字数据集合。杰奎琳·韦尼蒙特 (Jacqueline Wernimont) 解释说，在将死亡重新转化为数字的过程中，这些法案“创造了一个具有讽刺意味的理想化世界，在这个世界里，流行病和大规模死亡的报告看起来就像账簿一样干净有序。” 14从 18 世纪开始，欧洲人开始记录大量数据，并创建新的数学工具来检查这些数据，以加强政府、影响政策和说服人民。随着数字的积累加速，人类生活的越来越多的方面被记录成抽象的数字。从一开始，政府、教会和私人统计员就将越轨、死亡、犯罪和疾病的数字制成表格。新旧机构记录了生命和死亡过程的细节，当时和现在一样，违法的人留下了痕迹。从 18 世纪开始，统计学思维从根本上建立在对国家、人民以及常常被视为异常的人的数据收集的爆炸式增长之上。
Regularly published bills documenting the causes of deaths in parishes comprised some of the earliest recognizable collections of numerical data in seventeenth-century England. In recasting death into numbers, these bills, Jacqueline Wernimont explains, “produced an ironically idealized world in which the reporting of epidemic disease and mass death appeared as clean and orderly as an account book.”14 From the eighteenth century onward, Europeans dramatically began to record abundant data and create new mathematical tools to examine these data to strengthen governments, influence policy, and persuade their peoples. As the accumulation of numbers accelerated, more and more facets of human lives were recorded in abstract numerical terms. From the start, governments, churches, and private statisticians tabulated numbers about deviance, death, crime, and sickness. Institutions new and old recorded details about the course of life and death, and—then as now—people running afoul of the law left traces. Statistical thinking from the 1700s onward rested fundamentally on the explosion of the collection of data about states, their people, and, quite often, people deemed to be deviant.
最初，收集数字主要是描述性的，过程中很少进行计算或数学工作。伦敦统计学会于 1834 年成立，他们选择了一种印章，上面写着aliis exterendum （“由他人处理”）。他们寻求“简单地收集事实，让其他人”来解释它们。15然而，其他人则致力于开发新的工具来制作理解所有这些数字并以此为基础提出论点。16从18世纪开始，金融家、科学家和官僚们都开始开发新的数学和视觉手段来理解这些数据并据此提出主张，无论是说服投资者付钱还是影响政策。虽然我们的比利时天文学家凯特勒可能对后来的统计数据产生了最大的影响，但被称为“摄影师”的德国政治家、英国人口统计学家和金融家以及其他人都想出了使用新形式的数据驱动分析来重新制定治国方略和经济的方法。17 “统计学”一词已经大大脱离了其定性根源，一方面包含了从人到气候等所有事物的数据积累，主要是数字数据，另一方面包含了一套强大、迷人且经常被滥用的数学工具来得出结论和分析数据。18
This collecting of numbers was initially largely a descriptive affair, with scant calculation or mathematical work along the way. When the Statistical Society of London was founded in 1834, they chose a seal with the words aliis exterendum (“to be threshed out by others”). They sought “simply to gather the facts, leaving it to others” to interpret them.15 Others, however, worked to develop the new tools for making sense of all these numbers and making arguments based on them.16 Financiers, scientists, and bureaucrats alike began, from the eighteenth century onward, to develop new mathematical and visual means for making sense of this data and for making claims based on it, whether to convince investors to pony up or to affect policy. While our Belgian astronomer Quetelet may have had the greatest impact on subsequent statistics, German statesmen called “cameralists,” English demographers and financiers, and others devised ways to rework statecraft and economies using new forms of data-driven analysis.17 Having moved dramatically away from its qualitative roots, the term “statistics” came to incorporate, on the one hand, the accumulation of data, primarily numerical data, about everything from people to climate, and on the other hand, a set of powerful, beguiling, and often misused mathematical tools to draw conclusions and analyze data.18
正如 20 世纪 90 年代数据驱动的分析“颠覆”了当地杂货店的商品营销一样，对人口、产量和耕地面积的实证分析也挑战了旧的认知方式，以便进行统治。数据研究有可能取代其他形式的专业知识，从科学到车间再到药店。它不是对乡村的华丽描述，而是对动植物的计数。它不是对价值观的伦理讨论，而是试图定量模拟给定政策的影响。它不是对死亡的残酷现实，而是死亡率统计表。它不是对消费者潜在需求的专业知识，而是对每笔购买行为的收集和分析。它不是个别医生对药物的临床经验，而是衡量有效性和安全性的随机试验。它不是对申请大学的学生性格的判断，而是使用标准化测试来提供“客观”的测量。
Just as data-driven analysis “disrupted” the marketing of goods at your local grocery in the 1990s, empirical analysis of the population, production, and acres under tillage challenged older ways of knowing in order to rule. The study of data threatened to displace other forms of expertise, from science to shop floors to the drug store. Rather than lush descriptions of countryside, a counting of flora and fauna. Rather than an ethical discussion of values, an attempt to model the effects of a given policy quantitatively. Rather than the gruesome reality of death, tables of mortality statistics. Rather than expertise about potential desires of consumers, the collection and analysis of every purchase. Rather than the clinical experience of individual physicians with a drug, randomized trials to gauge effectiveness and safety. Rather than a judgment of the character of a student applying to college, the use of standardized tests to supply an “objective” measurement.
现代意义上的统计学的诞生源于人们意识到融合数据和数学分析可以为权力服务，但有时也可以制衡权力。
The birth of statistics in the modern sense comes from the realization that fusing data and mathematical analysis could serve power—but also, at times, could check power.
欢迎我们的比利时天文学家凯特勒 (Quetelet)。
Enter our Belgian astronomer, Quetelet.
天文学家观察社会世界
Astronomer Looking at the Social World
此后不久，对国家或地区的定量了解促使科学家尝试以完全不同的方式了解人类，改变我们理解自己的方式——作为道德人、物理存在和社会存在。政府和其他机构开始断断续续地收集有关死亡、犯罪和自杀的数据。他们中的大多数都严格掌握这些数据；许多甚至认为它们是国家机密。凯特勒试图获得这些数据，然后将其发表。他利用欧洲广泛的科学家网络从管理人员那里哄骗出数字，然后在自己的期刊上发表评分。19在互联网时代，期刊出版可能看起来慢得离谱——但这却是数据公开获取和流通的根本性转变。
Knowing states or countries quantitatively soon thereafter led scientists to try to know human beings quite differently, to change the way we understand ourselves—as moral people, as physical beings, as social beings. Governments and other institutions began collecting data on death, crime, and suicide in fits and starts. Most of them held this data closely; many even considered them secrets of state. Quetelet sought to obtain these numbers and then to publish them. He drew upon a broad European network of scientists to coax numbers from administrators and then published scores in his own journal.19 In the age of the internet, journal publication may seem ridiculously slow—but it was a radical transformation in the public availability and circulation of data.
他开始分析所有这些数据。他调整并简化了天文学家的数据分析方法，以发现人口数据中的规律性。一些规律性已为人所知一个多世纪——通常被当作天意组织世界的证据。历史学家凯文·唐纳利强调了凯特勒如何试图从死亡率等人类能动性有限的主题转向犯罪等“道德”领域，在这些领域，能动性至关重要。随着他获得有关犯罪和人类身体特征的数据，凯特勒开始认识到另一种规律性，即数据围绕平均值分组。
And he set to analyzing all this data. He adapted and simplified the data analysis of the astronomers, to find regularities in data on populations. A few regularities had been known for over a century—and were typically offered as evidence of Divine Providence organizing the world. The historian Kevin Donnelly has stressed how Quetelet sought to move from subjects like mortality, where human agency was limited, to “moral” domains like crime, where agency was paramount. As he gained access to data on crime and human physical characteristics, Quetelet came to recognize another kind of regularity, of the grouping of data around averages.
他并不满足于仅仅注意到这些规律。他立即赋予它们意义——以及一种现实形式。“同样的犯罪每年以同样的顺序出现，这种惊人的一致性，”凯特勒解释道，“对肇事者施以同等比例的同等惩罚，这是一个独特的事实，这要归功于法庭的统计数据。” 20这些统计定律似乎质疑了人类的自由意志。它们表明我们每个人都无法控制自己的命运。
He was not content simply to note these regularities. He immediately granted them significance—and a form of reality. “This remarkable constancy with which the same crimes appear annually in the same order,” Quetelet explained, “drawing down on their perpetrators the same punishments, in the same proportions, is a singular fact, which we owe to the statistics of the tribunals.”20 These statistical laws appeared to question human free will. They suggested that we do not each control our fate.
从错误理论到普通人
From Error Theory to Average Man
凯特勒认为，如果我们在大量人群中观察到“道德现象”，它们就会变得类似于物理现象。“观察到的个体数量越多，个体特征（无论是物理的还是道德的）就越被消除，而社会存在和保存的一般事实则占据了主导地位。” 21如何处理大量的个人观察？当时和现在一样，人们可以写很多小说，并希望从中得出一些永恒的人类状况。相反，凯特勒应用了一种新的数学技术，最初用于处理大量的天文观测。
Quetelet argued that if we observe “moral phenomena” in large numbers of people, they come to resemble physical phenomena. The “greater the number of individuals observed, the more do individual peculiarities, whether physical or moral, become effaced, and leave in a prominent point of view the general facts, by virtue of which society exists and is preserved.”21 How to deal with lots of individual observations? Then as now, one could write lots of novels, and hope to yield something of the eternal human condition. Instead, Quetelet applied a new mathematical technology originally for dealing with abundant astronomical observations.
为了建造天文台，凯特勒前往巴黎，在那里他了解到，通常由许多不同的人进行的大量天文观测可以转化为有关夜空中恒星和行星位置的相当确定的知识。如果几个人测量天空中某颗恒星的位置，观测到的位置会因时间、人与人之间以及仪器的不同而不同。
In his quest to build an observatory, Quetelet traveled to Paris where he learned how large numbers of astronomical observations, usually produced by many different people, could be converted into fairly certain knowledge about the positions of the stars and the planets in the night sky. If several people measure the position of a given star in the sky, the observed positions will vary from time to time and from person to person and from instrument to instrument.
伟大的数学家皮埃尔-西蒙·拉普拉斯和卡尔·高斯已经证明，对同一数量的多次天文观测往往符合我们通常所说的钟形或正态曲线。曲线的中心提供了最能得到证据支持的恒星体的位置。凯特勒利用这项天文技术做了一些新的事情，用于处理由不同组观测产生的大量数据眼睛。22他将这种推断方法应用到人类数据上——例如犯罪率、自杀率或人口身高。然后他做出了一个影响深远的飞跃——当时的科学并不完全支持这一飞跃。
The great mathematicians Pierre-Simon Laplace and Carl Gauss had shown that multiple astronomical observations of the same quantity tend to fall along what we often call the bell or normal curve. The center of the curve provides the location for a stellar body best supported by the evidence. Quetelet did something new with this astronomical technology for dealing with lots of data produced by different sets of eyes.22 He applied this way of inferring to data about human beings—data like the incidence of crime or suicide rate or heights of a population. And then he made an enormously consequential jump—one not entirely justified by the science of the time.
如果您和我连续数个夜晚对一颗恒星的位置进行大量观察，我们就会试图确定一个真实值：一颗恒星在天空中的位置。现在，如果我们测量一个陆军营所有成员的身高，我们可以轻松计算出平均身高。该平均值是从数据中抽象出来的，而不是试图测量某种真实的东西、某种外在的东西。这不像找到恒星的位置。
If you and I made a bunch of observations of the position of a star over many nights, we’d be trying to ascertain a real value: the position of one star in the sky. Now, if we measured all the heights of the members of an army battalion, we could easily compute the average height. That average would be an abstraction from the data, not an attempt to measure something real, something out there. It’s not like finding the position of the star.
凯特勒的天才之举——尽管它不够严谨——在于将人类的平均身高当作我们正在发现的真实量。他认为，一个群体的平均身高是一个真实的东西，就像恒星的位置一样——这个数字“客观地描述了这个群体”。23 尽管“数字在波动”，他写道，但我们知道“确实有一个数字，我们试图确定它的值，无论是个人的身高……还是北极星的赤经。” 24凯特勒坚持认为，正是这些数字描述了一个特定群体的“普通人”。
Quetelet’s flash of genius—whatever its lack of rigor— was to treat averages about human beings as if they were real quantities out there that we were discovering. He acted as if the average height of a population was a real thing, just like the position of a star—the number “objectively describes the population.”23 Despite “the fluctuation of numbers,” he wrote, we know “that there’s really a number whose value we seek to determine, whether it is the height of an individual . . . , or the right ascension of the polar star.”24 Just such numbers, Quetelet maintained, characterized the homme moyen or “average man” of a given population.
不管凯特勒所说的普通人现在听起来多么荒谬，制定能够代表整个人口的指标是我们政策的核心：犯罪率、GDP、智商。如果其中许多指标被理解为抽象概念，没有实际存在，那么其他指标，如不同种族或民族的先天智力，通常被解释为具有某种生物学现实。这种处理方式从根本上影响了教育和资源的获取，以及对人类差异性质的整个描述。我们将在接下来的章节中看到其他科学家如何在概念上制定指标凯特勒通过客观地描述“种族”而开辟了新的空间。
However ridiculous Quetelet’s average man sounds in retrospect, creating measures that characterize entire populations is central to our policy: crime rates, GDP, IQ. If many of these are understood as abstractions, having no real existence, others, like innate intelligence of different ethnic or racial groups, are often interpreted as having some biological reality. And this treatment has radically affected access to education and resources, as well as entire accounts of the nature of human difference. We’ll see in the chapters to come how other scientists created measures in the conceptual space that Quetelet opened up by characterizing a “race” objectively.
凯特勒专注于辨别“普通人”，这为描述特定社会（或 19 世纪习语中的“种族”）最典型的特征提供了一种工具。哲学家哈金解释说：“种族的特征在于其身体和道德品质的测量，这些品质总结在该种族的普通人身上。” 25以这种方式描述“种族”开辟了空间，可以比较“种族”之间、男女之间的差异，也可以进行发展分析，了解人类随时间推移的差异以及个人的发展。所有这些方面都是一门新的“人”科学的核心——一种科学地理解人性的新方法。
Quetelet’s focus on discerning the “average man” provided a tool for characterizing what is most characteristic of a given society—or “race,” in the nineteenth-century idiom. The philosopher Hacking explains, “A race would be characterized by its measurements of physical and moral qualities, summed up in the average man of that race.”25 Characterizing “races” in this way opened the space to understand comparative work of the differences between “races,” between men and women and also to undertake developmental analysis, to understand differences of human beings over time and well as the development of individual human beings. All these facets were central to a new science of “man”—a new approach to understanding human nature scientifically.
“人”的科学
The Science of “Man”
论文、传单、小说、淫秽诗歌——在十八世纪的欧洲启蒙运动中，所有这些都声称揭示了人类的真正本质。人类是否像托马斯·霍布斯和当今许多经济学家所说的那样，纯粹是自私的生物？他们有怜悯之心吗？他们从根本上是个人主义者还是社会主义者？他们积累了大量证据，其中大部分在我们看来都是轶事。正如凯特勒所认为的那样。“只有经验才能肯定地解决先验推理无法确定的问题，”他在关于辨别人类本质的文章中写道。“最重要的是，我们要保持对人类作为孤立、分离或个体状态存在的看法，只把人类视为物种的一部分。这样，抛开他的个体本性，我们就可以摆脱所有偶然因素，以及几乎不产生任何影响的个人特质对大众的影响，会自行消失，让观察者掌握总体结果。” 26
Treatises, leaflets, novels, bawdy poems—all claimed, across the European Enlightenment of the eighteenth century, to reveal the genuine nature of humanity. Were humans purely self-interested creatures, as Thomas Hobbes and many economists today might argue? Did they have pity? Were they fundamentally individualistic or social beings? They accrued scads of evidence, most of which appears anecdotal to us. As it did to Quetelet. “Experience alone can with certainty solve a problem which no a priori reasoning could determine,” he wrote about discerning the nature of humanity. “It is of primary importance to keep our view of man as he exists as an insulated, separate, or in an individual state, and to regard him only as a fraction of the species. In thus setting aside his individual nature, we get quit of all which is accidental, and the individual peculiarities, which exercise scarcely any influence over the mass, become effaced of their own accord, allowing the observer to seize the general results.”26
对于凯特勒来说，对人性的认识并非来自于对人类状况进行反思的空谈哲学家，也不是来自于捕捉个人生活细微差别的精心描述的现实主义小说。它来自于数学过程，这种数学过程将提取出真正人性的“普遍结果”，而不是这个或那个人的偶然事件。
For Quetelet, knowledge of human nature doesn’t come from armchair philosophers introspecting about the human condition or carefully described realistic novels capturing the nuances of an individual life. It will come from mathematical processes that will extract the “general results” characteristic of genuine human nature, not the accidents of this or that human being.
政府往往善于记录重大生命事件——出生、死亡以及国家和人民互动的时刻。在十九世纪初，就像现在一样，这些互动通常涉及警察、医生、教育工作者与他们所认为的犯罪或越轨行为作斗争。人们早就注意到出生和死亡数据的规律性。而凯特勒则强调了犯罪数据中的规律性。
Governments tend to be good at recording major life events—birth and death and those moments where states and people interact. In the early nineteenth century, as now, those interactions often involve police, doctors, educators contending with what they take to be crime or deviance. Regularities had long been noted in birth and death data. Quetelet, for his part, emphasized the regularities to be found in data about crime.
每年都以相同的顺序发生相同的罪行，对犯罪者施以相同的惩罚，比例也相同，这一惊人的一致性是一个奇特的事实，我们要归功于法庭的统计数据。在各种著作中，我尽最大努力将这一证据清楚地呈现在公众面前；我每年都会重申，我们以惊人的规律性支付了一笔预算——这是监狱、地牢和绞刑架的预算。27
This remarkable constancy with which the same crimes appear annually in the same order, drawing down on their perpetrators the same punishments, in the same proportions, is a singular fact, which we owe to the statistics of the tribunals. In various writings, I have done my utmost to put this evidence clearly before the public; I have never failed annually to repeat, that there is a budget which we pay with frightful regularity—it is that of prisons, dungeons, and scaffolds.27
即使在道德行为领域，数学规律也迅速出现。遵循当时许多最优秀的科学思想，凯特勒没有对犯罪的直接原因发表意见。他以典型的方式，把这些规律的证据当作证据存在着超越个人的某种东西，一种将人群统一起来的现实。他有争议地指出：“社会本身就包含着所有犯罪的萌芽，同时也包含着发展犯罪的必要设施。社会状态在某种程度上为这些犯罪做准备，而罪犯只是执行犯罪的工具。”要了解犯罪的速度和增长，我们需要了解社会的组织，而不仅仅是个人的组织。“每个社会状态都假设有一定数量和一定顺序的犯罪，这些只是其组织的必然结果。” 28在这里，凯特勒将物理学中必要原因的观点应用于社会世界。
Even in the domain of moral actions, mathematical regularities appeared quickly. Following much of the best scientific thinking of his day, Quetelet held back from opining on the precise immediate causes of crime. In his typical way, he took the evidence of these regularities as evidence of the existence of something above and beyond individual human beings, a reality unifying groups of people. “Society,” he controversially argued, “includes within itself the germs of all the crimes committed, and at the same time the necessary facilities for their development. It is the social state, in some measure, which prepares these crimes, and the criminal is merely the instrument to execute them.” To understand the pace and increase of crimes, we need to understand the organization of society, not just of individuals. “Every social state supposes, then, a certain number and a certain order of crimes, these being merely the necessary consequences of its organization.”28 Here Quetelet applied a vision of the necessary causes from physics to the social world.
事实上，他将自己的成就定义为表明通过数据观察到的道德现象与天文现象相似：“因此，通过这类研究，我们得出了一个基本原则：观察到的个体数量越多，个体的特殊性（无论是身体上的还是道德上的）就越被消除，而突出的是一般事实，社会正是凭借这些事实而存在和保存的。” 29理解人类社会意味着理解这些一般事实，而这通过积累越来越多的有关该社会及其人民的数据而成为可能。
Indeed, he characterized his own achievement as showing that moral phenomena, when observed through data, resembled astronomical phenomena: “we thus arrive in inquiries of this kind, at the fundamental principle: that the greater the number of individuals observed, the more do individual particularities, whether physical or moral, become effaced, and leave in a prominent point of view the general facts, by virtue of which society exists and is preserved.”29 Understanding human society meant understanding these general facts, something made possible by the accumulation of ever-larger amounts of data about that society and its people.
凯特勒坚持认为，他发现的代表不同社会的道德法则不应使希望黯淡，而应指向改善的可能性。他认为，犯罪源自社会组织，这应该是“安慰……通过展示改善人类的可能性，通过改变他们的制度、习惯、信息量，以及一般而言影响他们生存方式的所有因素。30
Quetelet insisted that his discovery of the moral laws characterizing different societies should not dim hope, but rather point toward the possibility of improvement. That crimes emerge from the organization of society, he argued, should be “consolatory. . . . by showing the possibility of ameliorating the human race, by modifying their institutions, their habits, the amount of their information, and, generally, all which influences their mode of existence.30
物化与客观性
Reification and Objectivity
伊恩·哈金认为，凯特勒极大地改变了我们理解世界的方式。通过他的工作，“仅描述大规模规律的统计定律”变成了“处理潜在真理和原因的社会和自然法则”。31寻找平均值不仅仅是对我们组成的群体的描述。对凯特勒来说，平均值是更真实的东西。平均值捕捉了存在于每个人之上的东西，关于群体本身的东西，关于每个群体成员的行为方式的东西。
Ian Hacking argues Quetelet dramatically transformed how we understand the world. With his work, “statistical laws that were merely descriptive of large-scale regularities” turned into “laws of society and nature that dealt in underlying truths and causes.”31 Finding the average isn’t just a description of a group that we make up. For Quetelet, that average is something more real. The average captures something existing above and beyond each person, something about the group itself, something about how each individual group member acts.
凯特勒将社会现象研究（往往属于正态曲线）与人类观察结果变化研究（也属于正态曲线）联系起来。正态曲线表征了人类潜在的变化性，并且让很多人惊讶的是，它表明了人类总体行为具有潜在的规律性。每一次自杀可能都是个人选择的产物，但一年中的所有自杀都属于可知的模式。将国家统计数据纳入正态分布有助于明确表征总体的属性。凯特勒与许多同时代思想家一起，使人们认为社会不仅仅是个人的集合。为捍卫个人的能动性和责任感，英国首相玛格丽特·撒切尔在 1987 年发表了著名的俏皮话：“只有个人的男人和女人……根本不存在社会” 32；然而凯特勒和他的继承者们却一再展示表征社会的数值定律。他称之为社会物理学。
Quetelet connected the study of social phenomena, which often fall into the normal curve, with the study of the variation of observations by human beings, which fall into the normal curve. The normal curve characterized underlying human variability, and, to the surprise of many, suggested an underlying lawlike behavior of humans in the aggregate. Every suicide may be a product of individual choice, but all the suicides in a year fall into knowable patterns. Putting state statistics into normal distributions helped make apparent properties characterizing the aggregate. Along with a number of other contemporaneous thinkers, Quetelet made the thinking of society as something more than merely a collection of individuals possible. In defense of individual agency and responsibility, Prime Minister Margaret Thatcher famously quipped in 1987, “There are individual men and women. . . . There is no such thing as society”32; yet Quetelet and his heirs repeatedly showed the numerical laws characterizing society. He called this social physics.
凯特勒重点关注普通人和社会的属性，他启发了其他科学去关注复杂的整体，即使我们不太了解这些整体的所有组成部分，我们也能够理解这些整体具有可理解的性质。历史学家西奥多·波特 (Theodore Porter) 认为，凯特勒 (Quetelet) 展示了“统计定律如何适用于大众，即使组成个体太多或太难以捉摸，以至于无法详细了解他们的行为。” 33早期数学与观测数据打交道的源头是物理学，19 世纪中叶，社会科学也接受了统计定律，认为它适合于理解自然世界。反过来，物理学的模型和统计程序又回到了社会科学。
With his focus on the attributes of the average man, of the social, Quetelet inspired other sciences to focus on complex wholes that can be understood to have intelligible qualities even if we little understand all their component parts. Historian Theodore Porter argues that Quetelet showed how “statistical laws can prevail for a mass even when the constituent individuals are too numerous or too inscrutable for their actions to be understood in any detail.”33 The source of much of the early mathematics for contending with observational data, physics, followed in the middle years of the 1800s the social sciences in embracing statistical law as appropriate for understanding the natural world. In turn, models and statistical procedures from physics came back to the social sciences.
阻止革命
Heading Off Revolution
凯特勒式的社会物理学并非简单地描述。它规定了——告诉人们该做什么，用强有力的道德语言描述了现代社会在工业化和殖民化过程中所必须的改进。“人类之友”必须研究社会中缓慢的统计变化，以便追求所期望的逐步修改。34革命和混乱并不存在。西奥多·波特评论说，社会物理学必须“被视为对渐进自由主义精神的社会秩序的赞歌”。35如果钟形曲线的中心区域是正常的，那么它的外围显然是病态的。了解人群是试图照顾和改善社会环境以类似法律的方式造成的越轨或病态个体的先决条件。从这个角度来看，越轨承担了新的角色。波特解释说：“凯特勒对中庸的理想化意味着，所有偏离中庸的行为都应该被视为有缺陷的，是错误的产物。” 36这个错误可以通过科学来认识。
Social physics in the manner of Quetelet didn’t simply describe. It prescribed—told what to do, in a potent moral language of the improvement necessary required for modern societies engaged in industrialization and colonization. The “friends of humanity” must study the slow statistical transformations in society in order to pursue the gradual modifications desired.34 Revolutions and disruptions need not apply. Theodore Porter remarks, social physics must “be recognized as a paean to social order in the spirit of gradualist liberalism.”35 If the central region of the bell curve was normal, its peripheries were evidently the realm of the pathological. Understanding populations was the prerequisite for attempting to take care of—and improve—deviant or pathological individuals caused in a lawlike way by their social settings. In this view, deviance took on a new role. Porter explains, “The implication of Quetelet’s idealizations of the mean was that all deviations from it should be regarded as flawed, the product of error.”36 And this error could be known through science.
伊恩·哈金在其杰出著作《驯服机遇》中指出，“普通人带来了一种关于人口的新信息，也带来了一种关于如何37凯特勒的继承者们抓住了他为描述人和种族所做的努力，并将他们推向了更高的境界。凯特勒试图通过统计数据来改善人类种族。他的努力表明了用新的“客观”测量方法来描述人类的权力，以及改善社会“种族平均素质”的必要性。他的思想的继承者们深化了这些思想，并以此为基础，创造了一种新的科学种族主义，这种种族主义不同于凯特勒的自由主义改进，但深受其惠。关键人物是查尔斯·达尔文的表弟弗朗西斯·高尔顿爵士，他将凯特勒的工作变成了一门新的“个体差异科学” 。38
In his remarkable The Taming of Chance, Ian Hacking argues that “the average man led to both a new kind of information about population and a new conception of how to control them.”37 The heirs of Quetelet seized on his efforts to characterize people and races—and pushed them much further. Through statistics, Quetelet sought to improve human races. His efforts suggested the power of characterizing humans with new “objective” measurements and the need for improving social “average qualities of race.” The heirs to his thought deepened and ran with these to craft a new scientific racism distinct from Quetelet’s liberal improvement yet deeply indebted to it. The key figure was a cousin of Charles Darwin, one Sir Francis Galton, who turned Quetelet’s work into a new “science of individual differences.”38
迈向人类进步的新科学
Toward New Sciences of Human Improvement
1891 年 2 月，凯特勒的英国支持者弗洛伦斯·南丁格尔在写给高尔顿的一封信中列出了一系列紧迫的政策问题。改革派和他们的反对者长期以来一直在没有确凿证据的情况下争论政策。她要求用统计数据来指导当权者，回答以下紧迫问题：
In February 1891, that British advocate of Quetelet, Florence Nightingale, penned a long list of pressing policy questions in a letter to Galton. Reformers and their opponents had long debated policy without solid evidence. She asked for statistics to inform power, to answer such pressing issues as:
法律惩罚的效果——即监禁对犯罪产生的威慑作用或者鼓励作用。
The results of legal punishments—i.e., the deterrent or encouraging effects upon crime of being in gaol.
...
...
教育对犯罪有何影响？
What effect has education on crime?
谈到印度，她问道：
Speaking of India, she asked:
那里的人民是富裕了还是贫穷了，吃得好穿得坏了还是变坏了？他们的体力是否在衰退？39
Whether the peoples there are growing richer or poorer, better or worse fed and clothed? Whether their physical powers are deteriorating or not?39
我们应该使用数据和统计分析来回答这些问题的想法在今天看来已经很平庸了。我们必须明白这些知识形式如何成为显而易见的统治和权力工具。同时代人对此并不了解。她呼吁高尔顿“写下他希望统计的其他重要领域，并教导人们如何使用这些统计数据来制定法律，更精确、更有经验地管理我们的国家生活。” 40高尔顿的目标甚至更高，正如我们将看到的，他要管理的正是全球不同种族的智力和身体素质。
The idea that we ought to use data and statistical analysis to answer such questions seems banal today. We have to understand how such forms of knowledge became obvious tools of governance, of power. It wasn’t obvious to contemporaries. She called on Galton to “jot down other great branches upon which he would wish for statistics, and for some teaching how to use these statistics in order to legislate for and to administer our national life with more precision and experience.”40 Galton aimed, as we’ll see, even higher, to administer nothing less than the intellectual and physical qualities of diverse races of the globe.
1915年，一位年轻的剑桥毕业生登上船返回印度。随行的还有当时最具活力、最令人兴奋的期刊，该期刊致力于通过数据了解现状和构建更美好未来的科学。这本名为《Biometrika》的期刊体现了一种新的数据驱动方法，用于解决生物和社会问题，包括种族和遗传问题——这极大地实现了凯特勒的梦想。毕业生普拉桑塔·C·马哈拉诺比斯认为，这些技术无论是殖民地印度还是未来独立的印度都可以采用，以统计的方式更好地了解自己，从而指导其经济和社会发展。作为印度统计学的伟大制度化者，马哈拉诺比斯改变了这些方法，但有时也会缓和它们的傲慢。马哈拉诺比斯带来的新科学源于数据、达尔文主义和对大英帝国的信任危机的交汇。这些科学产生了数理统计学，它诞生于英国强权之下——同时也引发了人们对颓废和衰落的巨大担忧。
In 1915, a young graduate from Cambridge boarded a ship to return home to India. With him was a complete run of the most dynamic, the most exciting periodical of the day, focused on the science for understanding the present and building better futures through data. The journal, Biometrika, embodied a new data-driven approach to biological and social problems, including questions of race and heredity—a great furthering of the dreams of Quetelet. The graduate, Prasanta C. Mahalanobis, saw in them technologies both colonial and a future independent India could adapt, to know itself better in statistical ways, to guide its economic and social development. As the great institutionalizer of statistics in India, Mahalanobis transformed these methods while tempering at times their hubris. The new sciences Mahalanobis brought with him emerged from the intersection of data, Darwinism, and a crisis of confidence in imperial Britain. These sciences produced mathematical statistics, born amid British power—and great fears of decadence and decline.
道德恐慌可以创造新的科学。十九世纪末英国精英阶层对帝国的衰落深感忧虑。弗洛伦斯·南丁格尔在 1858 年写道：“似乎是时候了，只有英国民族才能维护帝国的完整性。” 1精英阶层的出生率下降、“人口有限”、酗酒和海外失败都表明帝国陷入了危机。优生学这一新领域及其支持的数据不仅提供了一种诊断社会的方法，而且提供了一种尝试治愈社会的方法——一种让英国再次伟大的方法。
Moral panics can create new sciences. Late nineteenth-century elite Britons were consumed with worries about their empire’s degeneracy. “The time appears to have arrived,” wrote Florence Nightingale in 1858, “when by the British race alone must the integrity of that Empire be upheld.”1 Declining birthrates among the elite, “limited population,” alcoholism, and failures abroad all spoke to an empire in crisis. A new field, eugenics, and the statistics to support it provided one way not only to diagnose society, but to attempt to cure it—a way to make Britain great again.
现代时代的危机
A Crisis of the Modern Era
当绅士学者弗朗西斯·高尔顿考察维多利亚时代的英国同胞时，他发现他们存在不足：“我们需要更有能力的指挥官、政治家、思想家、发明家和艺术家，”他在一篇名为《遗传的天赋和性格》的文章中写道。“我们种族的自然资质并不比半野蛮时代更好”，尽管“我们出生的环境比过去复杂得多”。现代文明太过分了。“当今最杰出的人才似乎在智力负担过重的情况下步履蹒跚、停滞不前。” 2需要天才，但天才却供不应求。教育永远不够，因为根本没有足够的有天赋的男人和女人，没有足够的天生的天才来应对时代的复杂性。英国需要更多的天才，更多才华横溢的人。高尔顿认为，他们需要被培养。
When the gentleman-scholar Francis Galton inspected his fellow men of Victorian Britain, he found them wanting: “We want abler commanders, statesmen, thinkers, inventors, and artists,” he wrote in an article called “Hereditary Talent and Character.” “The natural qualifications of our race are no greater than they used to be in semi-barbarous times,” even though “the conditions amid which we are born are vastly more complex than of old.” Modern civilization was all too much. “The foremost minds of the present day seem to stagger and halt under an intellectual load too heavy for their powers.”2 Genius was needed, but was in short supply. Education would never be enough, for there were simply not enough gifted men and women, not enough born geniuses, to confront the complexity of the times. England needed more geniuses, more people of extraordinary talent. They needed, Galton decided, to be bred.
高尔顿的堂兄查尔斯·达尔文虽然名声显赫，但名声不佳，他的著作为我们指明了前进的方向。3在《物种起源》中，达尔文以人类饲养家畜（如展示鸽和纯种狗）为例，阐述了他对自然选择的解释。正如人类饲养者选择他们想要的特征一样，动物的某些特征也是随着时间的推移，在特定的环境中，物种逐渐被选择。在高尔顿看来，人类低估了自己影响物种的能力。高尔顿解释说：“人类对动物生命的权力是巨大的。后代的身体结构似乎和粘土一样具有可塑性，完全由饲养者的意志控制。”不仅仅是身体特征，心智也可以改变：“我希望比我所知的更明确地表明，心理品质同样受到控制。” 4
The writings of Galton’s illustrious if infamous cousin, Charles Darwin, offered a way forward.3 In his Origin of Species, Darwin used the example of human breeding of domestic animals, such as show pigeons and pedigree dogs, to motivate his account of natural selection. Just as human breeders select features they desire, certain features of animals are selected for, as it were, over time in particular environmental niches. Humans, in Galton’s estimation, underestimated their own power to affect their species. “The power of man over animal life,” Galton explained, “is enormously great. It would seem as though the physical structure of future generations was almost as plastic as clay, under the control of the breeder’s will.” Not just physical traits, but equally the mind could be altered: “It is my desire to show more pointedly than—so far as I am aware—has been attempted before, that mental qualities are equally under control.”4
高尔顿很快创造了“优生学”一词，用来描述有意识地改善人类素质（尤其是人类的民族种族）的努力。优生学很快成为欧洲、美国和世界许多左翼和右翼政治计划的核心。高尔顿本质上是种族主义者，但他的主要关注点是阶级。他的建议往往是异想天开的，尤其是与英国以外的优生计划相关的强制绝育和种族灭绝相比：
Galton soon coined the term “eugenics” to describe the conscious effort to improve the quality of human beings— and national races of human beings in particular. Eugenics quickly became central to many left- and right-wing political programs across Europe, the United States, and the world. Racist to its core, to be sure, Galton’s primary focus nevertheless was class. His suggestions often were whimsical, especially in comparison with the forced sterilizations and genocides associated with eugenic programs to come outside of Britain:
那么，让我们放飞想象，想象一个乌托邦——或者说拉普达，如果你愿意的话——那里的女孩和青年的竞争性考试制度已经发展到涵盖了身心所有重要素质的程度，而且每年有一笔可观的资金用于资助那些许诺生育出的孩子长大后能成为国家杰出公仆的婚姻。5
Let us, then, give reins to our fancy, and imagine a Utopia—or a Laputa, if you will—in which a system of competitive examination for girls, as well as for youths, had been so developed as to embrace every important quality of mind and body, and where a considerable sum was yearly allotted to the endowment of such marriages as promised to yield children who would grow into eminent servants of the State.5
与他那个时代的许多哲学家和经济学家不同，高尔顿从根本上反对平等主义。6 “我反对自然平等的主张，”他写道。“我没有耐心偶尔会有人提出这样的假设……婴儿出生时几乎都一样，而造成男孩与男孩、男人与男人之间差异的唯一因素是坚持不懈的努力和道德努力。” 7高尔顿坚持认为，并不是所有人都生而平等，并不是所有的市场主体都具有相当的智力。在他看来，自由主义政治思想和自由主义经济学是错误的。
Unlike many philosophers and economists of his time, Galton was fundamentally anti-egalitarian.6 “I object to pretensions of natural equality,” he wrote. “I have no patience with the hypothesis occasionally expressed. . . . that babies are born pretty much alike, and that the sole agencies in creating differences between boy and boy, and man and man, are steady application and moral effort.”7 All people were not created equal, Galton insisted, and not all market agents had comparable mental capacities. Liberal political thought and liberal economics were just wrong in his estimation.
我们将优生学和科学种族主义与极右翼、纳粹联系在一起。但 1900 年左右的情况并非如此。直到第二次世界大战，许多进步人士和保守派都认为科学能够通过改善人类种族来改善人类的命运；事实上，一位支持者指出，对优生学的信仰提供了“一个人视野广度和对我们种族未来的无私关怀的完美指标”。8统计科学将用基于证据的人类改进新科学取代旧的偏见：对自然人类等级制度的描述可以轻松地从描述转变为处方。9
We associate eugenics and scientific racism with the far right, with Nazis. Things were otherwise around 1900. Many progressives as well as conservatives up to World War II saw science as capable of improving the human lot by improving the human race; indeed, one proponent noted, belief in eugenics offered “a perfect index of one’s breadth of outlook and unselfish concern for the future of our race.”8 Statistical sciences were to replace the bigotries of old with evidence-based new sciences of human improvement: accounts of natural human hierarchies that moved easily from description to prescription.9
为了改良人类，高尔顿需要探索天赋和人类卓越品质的源泉。后天培养并不是原因。高尔顿利用伟人的传记词典，开始研究家族中天赋和天才的密度。在他 1869 年的长期研究《遗传天才》中，高尔顿研究了杰出家族，并将历史国家与世界各地的国家进行了比较。尽管他的案例数量众多，但他的方法却是直觉和轶事。他通常认为现代人都比古希腊人低劣，非欧洲人——他称之为“种族”——比欧洲人低劣。
To improve the species, Galton needed to explore the wellsprings of talent and human excellence. Nurture was no explanation. Using biographical dictionaries of great men and women, Galton began investigating the density of talent and genius within families. In his long study Hereditary Genius of 1869, Galton studied prominent families and compared historical states with those around the earth. Despite the large number of his cases, his approach was intuitive and anecdotal. He generally argued that modern peoples were all lesser than ancient Greeks and that non-European peoples— he called them “races”—lesser than European ones.
虽然书中的方法主要是轶事，但高尔顿利用了凯特勒的正态曲线来支持他对人和种族进行排名的新想法。凯特勒使用正态曲线来理解整个群体的品质。高尔顿用同样的曲线来理解一个群体内部的变化群体。凯特勒可能在寻找英国人的平均身高。高尔顿则试图了解身高的极端值。他追求的是天赋，而不是身高，但他将同样的工具应用于两者。法国社会学家阿兰·德罗西埃解释说，高尔顿将正态曲线用作“一种允许对个体进行分类的偏差定律，而不是误差定律”。10天文学家认为需要消除的误差，高尔顿则认为需要对个体进行排名和分类。每个获得列出其百分位表现的考试成绩的孩子都生活在高尔顿帮助创造的世界中。
While the approach in his book is largely anecdotal, Galton drew on Quetelet’s normal curve to support his new ideas of ranking people and races. Quetelet used the normal curve to understand the qualities of a group as a whole. Galton used the same curve to understand variation within a group. Quetelet might seek the mean stature of Englishmen. Galton sought to understand the extremes of stature. His quarry was talent, not height, but he applied the same tools to one and other. The French sociologist Alain Desrosières explains, Galton used the normal curve as “a law of deviation allowing individuals to be classified, rather than as a law of errors.”10 What astronomers saw as errors to be eliminated, Galton saw as individuals to be ranked and classified. Every child getting test scores listing their percentile performance lives in the world Galton helped create.
然而，在对名门望族优秀品质的研究中，有一个主要的症结。身高极高的人生出高个子的孩子，但平均而言，这些孩子不如父母高，回归到人口平均身高。类似的观察结果描述了人类和动物的广泛特征。对于动物饲养者（无论是人类还是其他动物）来说，这是一个难题，它将限制人们培育所谓优秀人类的尝试。如何理解它？答案来自高尔顿对凯特勒对正态曲线的热情应用的重新研究。
And yet there was a major sticking point to all this investigation of excellence in distinguished families. Extremely tall people had tall children, but, on the average, those children were not as tall as their parents, reverting toward a population average height. Similar observations describe a wide range of human and animal traits. For breeders of animals—human or otherwise—this was a puzzle, one that would limit attempts to breed supposedly superior human beings. How to understand it? The answer would come from Galton’s reworking of Quetelet’s enthusiastic applications of the normal curve.
为什么高个子父母的后代往往不如父母高？更普遍地说，为什么一个人类群体的属性随着时间的推移几乎保持不变？高尔顿通过他所谓的“回归”来解释这两种现象，从数学上捕捉到“理想的平均子女类型偏离父母类型的趋势，‘回归’到可以粗略地、也许公平地描述为平均祖先类型”。11通过他的统计调查，他发现后代的回归量与其父母偏离平均值的程度之间存在强大的数学关系。他不仅表明这种关系是线性的，而且还进行了我们今天所说的（多亏了高尔顿）应用于平均子女类型的线性回归。数据，找到简单线性方程的系数，如y = mx+b。
Why do offspring of tall parents tend not to be as tall as their parents, and more generally why do the attributes of a human group stay nearly constant over time? Galton came to explain both phenomena through what he called “regression,” mathematically capturing the “tendency of that ideal mean filial type to depart from the parent type, ‘reverting’ towards what may be roughly and perhaps fairly be described as the average ancestral type.”11 With his statistical investigation, he discovered a powerful mathematical relationship between the amount of reversion of offspring and the extent of their parents’ deviation from the mean. He not only showed the relationship to be linear, but also undertook what we would call today, thanks to Galton, the linear regression applied to the data, finding the coefficients of a simple linear equation like y = mx+b.
弗朗西斯·高尔顿。“遗传地位向平庸回归。” 《英国及爱尔兰人类学研究所期刊》第 15 卷（1886 年）：246–63 页。图 IX。
Francis Galton. “Regression Towards Mediocrity in Hereditary Stature.” The Journal of the Anthropological Institute of Great Britain and Ireland 15 (1886): 246–63. Plate IX.
高尔顿正在对生育过程的各个方面进行建模，因此他最初对回归的研究只涉及将父母身高作为 x，将子女身高作为 y，因为他正在研究一个单向的生物过程。但他很快意识到，他的回归过程可以脱离其生物学基础，并用于大量数据。在研究“回归”过程时，高尔顿不知不觉地想到了一个更广泛的概念，即统计回归。
Galton was modeling facets of the process of generation, so his initial work with reversion involved only treating the parental heights as the x’s and only the children’s heights as the y’s, for he was looking at a unidirectional biological process. But he soon realized that his process of regression could be detached from its biological mooring and used on a vast array of data. In investigating the process of “reversion,” Galton had unknowingly hit on a much broader concept, namely that of statistical regression.
相关性和数据
Correlation and Data
高尔顿所做的远不止引入一种强大的新方法来建模数据并根据数据进行预测。凯特勒研究社会。高尔顿研究的是分布中的个体。他希望有更好的技术来了解和排列个体以及种族。在研究父母和子女身高等属性对之间的关系时，高尔顿还引入了“相关性”，或者我们现在称之为相关性。
Galton did far more than introduce a powerful new approach to modeling data and making predictions from data. Quetelet studied society. Galton studied individuals in a distribution. He wanted better techniques to know and rank individuals and to know and rank races. In studying relations between pairs of attributes, such as the height of parent and child, Galton also introduced “co-relation,” or as we would now term it: the correlation.
虽然各国政府都在制作越来越多的统计数据，但却未能积累足够多高尔顿最感兴趣的数据——对大量人口的“主要身体特征”进行详细调查，例如“视觉敏锐度、色觉、眼睛的判断力、听觉、最高音调、呼吸能力、拉挤强度、出拳速度、臂展、身高、站姿和坐姿以及体重”。12收集这些数据非常困难，因此高尔顿在 1884 年南肯辛顿国际卫生博览会上设立了一个人体测量实验室。实验室用 17 种方法对 9,337 人进行了测量。他解释说“定期测量”将有助于家庭追踪个人发展情况，并“发现整个国家及其各个部分的效率”。这样的记录“使我们能够比较学校、职业、住所、种族等”。13产生的数据一直持续到 20 世纪才被研究。心理学史学家库尔特·丹齐格 (Kurt Danziger) 解释说，高尔顿的人体测量学“将个人表现定义为先天生物因素的表现，从而将其与任何可能的社会影响隔离开来。” 14
While governments were producing an ever-i ncreasing number of statistics, they failed to accumulate enough of the data that most interested Galton—detailed investigations of the “chief physical characteristics” of a wide selection of the population, qualities such as “Keenness of Sight; Colour-Sense; Judgment of Eye; Hearing; Highest Audible Note; Breathing Power; Strength of Pull and Squeeze; Swiftness of Blow; Span of Arms; Height, standing and sitting; and Weight.”12 So challenging was collecting this data that Galton set up an Anthropometric Laboratory at the International Health Exhibition of 1884 in South Kensington. The laboratory measured 9,337 people in seventeen ways. He explained that “periodical measurements” would be useful to families in tracking their individual development, and to “discover the efficiency of the nation as a whole and in its several parts.” Such records “enable us to compare, schools, occupations, residences, races, &c.”13 The data produced would continue to be studied well into the twentieth century. Galton’s anthropometry, historian of psychology Kurt Danziger explains, “defin[ed] individual performances as an expression of innate biological factors, thereby sealing them off from any possibility of social influence.”14
高尔顿的风格为理解人类差异提供了一种全新的方法。根据凯特勒的理论，数据分析可以揭示可量化的人类行为和属性的共性和范围。根据高尔顿的理论，每个人都可以被置于这些范围内并进行排名：前 5%，后 10%。受 例如，高尔顿通过观察大量人类，发明了心理测试，试图将每个人置于人类能力的可测量范围内。随后，以统计学方式研究大量“受试者”的整个科学也应运而生。历史学家丹齐格解释说：“高尔顿及其思想继承者卡尔·皮尔逊的工作使一种证明心理知识主张的新方法成为可能。”“要对个人做出有趣而有用的陈述，没有必要让他们接受密集的实验或临床探索。只需要将他们的表现与其他人的表现进行比较，在某个个人表现的集合中为他们分配一个位置。” 15这种方法很快就成为一门大生意。当高尔顿这样的先驱者努力获取足够规模的数据时，人们对此类调查的兴趣很快便日渐高涨，尤其是在第一次世界大战后的美国。16
Galton’s style enabled a dramatic new approach to understanding human differences. Following Quetelet, analysis of data could reveal the commonalities and range of quantifiable human behavior and attributes. And following Galton, each individual could be placed and ranked within those ranges: the top 5 percent, the bottom 10 percent. Inspired by Galton’s work in observing large numbers of human beings, mental tests, for example, emerged from the effort to place each person amid the range of measured human capacities. And entire sciences of examining large numbers of “subjects” in statistical ways emerged in its wake. “A new method for justifying psychological knowledge claims had become feasible” with the work of Galton and his intellectual successor Karl Pearson, explains the historian Danziger. “To make interesting and useful statements about individuals it was not necessary to subject them to intensive experimental or clinical exploration. It was only necessary to compare their performance with that of others, to assign them a place in some aggregate of individual performances.”15 And it didn’t take long for an approach to become big business. While pioneers like Galton struggled to get data at an adequate scale, a vast appetite for such inquiries would soon open, especially in the United States after the First World War.16
弗朗西斯·高尔顿 (Francis Galton)，《人体测量实验室；由弗朗西斯·高尔顿 (Francis Galton)，FRS 编纂，用于测定身高、体重、跨度、呼吸能力、拉挤强度、吹气速度、听觉、视觉、色觉和其他个人数据》 (伦敦：威廉·克劳斯 (William Clowes)，1884 年)，第 13 页。
Francis Galton, Anthropometric Laboratory; Arranged by Francis Galton, FRS, for the Determination of Height, Weight, Span, Breathing Power, Strength of Pull and Squeeze, Quickness of Blow, Hearing, Seeing, Colour-Sense, and Other Personal Data (London: William Clowes, 1884), 13.
最重要的是，高尔顿揭示了如何通过调查大量人群来识别和锁定个人。大量关于大量人群的数据可以让科学家、营销人员、军队和间谍更好地了解你，并锁定你。我们生活在这样一个世界里，我们的个性被量化，以互联网上所有其他用户的参考为参考，广告投放算法利用这种差异量化来争夺我们的注意力。
Above all, Galton revealed how surveying a mass of people makes recognizing—and targeting—the individual possible. Lots of data about lots of people allows scientists, marketers, militaries, spies to better know you—and target you. We live in such a world, where our individuality is quantified in reference to all other users of the internet, and where ad-serving algorithms exploit this quantification of difference to compete for our attention.
生物识别技术制度化
Institutionalizing Biometrics
不知疲倦的高尔顿本人并没有将他的新统计方法制度化。他同样不具备使之严谨的数学技能。他的继承人卡尔·皮尔逊借鉴了高尔顿的思想和资金支持，在两个方面都进行了研究。作为贵格会教徒、自由思想家、数学家、社会主义者、女权主义者和优生学家，皮尔逊有一个“宏伟的愿景，即创造一种统计生物学作为有效优生学的基础，同时发展一种可以应用于几乎所有人类知识领域的数理统计学”——用他的传记作者西奥多·波特的话来说。17卓越的数理统计学将使其扩展到凯特勒梦想的所有现象，即整个社会改革领域。18在高尔顿和英国布商会等赞助人的帮助下，他的领域建设使优生学和一种专横的新统计方法制度化，用于社会和政治计划。
The indefatigable Galton did not himself institutionalize his new statistical approach. He likewise did not have the mathematical skills to make it rigorous. Drawing on Galton’s ideas and financial support, his intellectual heir, Karl Pearson, worked at both. A descendant of Quakers, freethinker, mathematician, socialist, feminist, and eugenicist, Pearson had a “grand vision, the creation of a statistical biology as the basis of effective eugenics and, concomitantly, the development of a mathematical statistics that could be applied to virtually all areas of human knowledge”—in the words of his biographer Theodore Porter.17 Superior mathematical statistics would enable its expansion to the whole range of phenomena Quetelet dreamed of, an entire spectrum of social reform.18 His field-building institutionalized eugenics and an imperious new statistical approach to social and political programs, with the help of patrons such as Galton and, of all things, the Worshipful Company of Drapers.
为了完成所有这些工作，皮尔逊需要数据、处理数据的劳动力和新的数学。19正如他在一次著名的演讲中指出的那样，“这项工作本质上是一项合作调查的结果，调查范围涉及许多年，并依靠一群合作者”，他们制作并分析了“我的结果完全依赖于的大量数据”。20皮尔逊与一群员工辛苦工作了几十年，才将他的项目结出硕果；一代伟大的统计学家在他的指导下与他一起工作，改变了我们使用数据的方式。皮尔逊管理多个实验室，包括独立的生物统计和优生实验室，项目、方法、人员和资金各不相同。21在两位女助手爱丽丝·李 (Alice Lee) 和埃塞尔·埃尔德顿 (Ethel Elderton) 的帮助下，他收集了广泛的数据用于广泛的统计调查，并发表了基于这些数据的结果，主要是在他创办和运营的期刊上。
To do all this Pearson required data, labor to process that data, and new mathematics.19 As he noted while giving a prestigious lecture, “the work is essentially the result of a co-operative investigation extending over a number of years, and depending upon a body of collaborators” who produced and analyzed “the extensive data on which my results entirely depend.”20 Pearson toiled for decades with a cadre of workers to bring his projects to fruition; a generation of great statisticians worked under and with him and changed how we all use data. Pearson ran multiple laboratories, including separate biometric and eugenic laboratories, with distinct projects, methods, staff, and funding.21 With the help of two women assistants in particular, Alice Lee and Ethel Elderton, he amassed a wide range of data for a wide range of statistical investigations, and published results based on them, mostly in journals he founded and ran.
获取数据是一项艰苦的工作。1903 年，伦敦发现了一个瘟疫坑。不到一周后，“我的一名工人 SM Jacob 先生以不同寻常的热情‘恳求’所有头骨和骨骼”来获取 Pearson 的研究成果。22大多数数据获取都比较平淡无奇。为了扩展 Galton 对身体和智力遗传的研究，Pearson 和他的团队在校长和教师阅读的杂志上发布请求，要求他们记录对兄弟姐妹的大量观察结果并对他们进行智力排名。他们发出了 6,000 份表格，并从众多学校收回了约 4,000 份（见图）。 “绝对分类和制表是一项艰巨的工作，”皮尔逊解释道，并感谢了一支杰出的女性团队：“爱丽丝·李小姐，理学博士；玛丽·莱文茨小姐，文学硕士，E.佩兰小姐，玛丽·比顿小姐和玛格丽特·诺卡特小姐”，然后指出“计算的主要工作落在了爱丽丝·李博士身上。” 23
Getting data was hard work. In 1903, a plague pit was opened in London. Less than a week later, “one of my workers, Mr. S. M. Jacob, had with unwonted energy ‘begged’ the whole of the crania & skeletons” for Pearson’s work.22 Most data acquisition was more prosaic. Aiming to extend Galton’s studies of the inheritance of physical and mental capacities, Pearson and his team placed requests in magazines read by headmasters and teachers asking them to record a multitude of observations on pairs of siblings and to rank them intellectually. They sent out 6,000 forms and got back some 4,000 from a wide range of schools (see illustration). “The absolute classification and tabling has been a work of great labour,” Pearson explained, thanking a team of exceptional women: “Miss Alice Lee, D.Sc.; Miss Marie Lewenz, M.A., Miss E. Perrin, Miss Mary Beeton and Miss Margaret Notcutt” before noting that the “chief labour of computing has fallen upon Dr Alice Lee.”23
即使有了新机器，处理数据仍然非常困难，而且成本高昂。高尔顿支持优生学实验室，1903 年，英国布料商同业公会向皮尔逊提供了 500 英镑，用于他的生物统计实验室，这让他开始向艾丽丝·李支付报酬，此前艾丽丝·李曾自愿为他进行大量计算，并与他合作。“她的职责包括减少数据、计算相关系数、创建条形图……和计算一种新统计数据”——卡方——以及监督男性和女性计算器。24机器计算成为皮尔逊实验室工作的核心，以至于一位访客注意到“对掌握计算细节的专注”可能会掩盖新的数理统计学。25这些工作的大部分导致了大量打印表格的收集。今天很难认识到这些表格作为计算基础设施对于数理统计学的发展有多么重要。
Processing data was arduous and expensive, even with the help of new machines. Galton supported the Eugenics Lab, and in 1903, the Worshipful Company of Drapers granted Pearson £500 for his Biometric Laboratory, which allowed him to begin paying Alice Lee, who had previously undertaken extensive calculations for him on a volunteer basis, as well as collaborating with him. “Her duties included reducing data, computing correlation coefficients, creating bar charts . . . and calculating a new kind of statistic”— chi-squared—as well as supervising calculators male and female.24 Calculation with machines became so central to the work at Pearson’s laboratory that one visitor noted a “preoccupation with mastery of details of calculation” that could obscure the new mathematical statistics.25 Much of this labor resulted in major collections of printed tables. It is hard to appreciate today just how essential such tables were as computational infrastructure for the growth of mathematical statistics.
《遗传调查数据论文》，收录于卡尔·皮尔逊的《论人类遗传规律：II。论人类心智和道德特征的遗传及其与身体特征遗传的比较》。《Biometrika 》第 3 卷，第 2/3 期（1904 年）：131–90，第 163 页。
Data Paper for Heredity Investigations, in Karl Pearson, “On the Laws of Inheritance in Man: II. On the Inheritance of the Mental and Moral Characters in Man, and Its Comparison with the Inheritance of the Physical Characters.” Biometrika 3, no. 2/3 (1904): 131–90, at p. 163.
当他的女同事们经常专注于乏味的计算时，皮尔逊也鼓励她们从事更高级的工作，并经常与她们一起发表文章。例如，他主张“埃尔德顿小姐不应再被称为办事员，而应成为弗朗西斯·高尔顿学者。她完全有能力做原创工作。”除了对统计学的贡献外，她们还可以成为当地社会工作的领导者。“最理想的是，在优生实验室接受过培训的人应该从事某种公共或市政服务工作，比如处理精神缺陷者或残疾儿童等。这样，我们将发展成为一所实践优生工作的培训学校。” 26这些女性中最著名的是 FN (弗洛伦斯·南丁格尔)·大卫 (以著名的健康改革家命名)，她在统计学领域拥有传奇般的职业生涯，包括担任加州的教授。
While his women co-workers were often preoccupied with the tedium of calculating, Pearson also encouraged their higher-level work and often published with them. He argued, for example, “that Miss Elderton be no longer spoken of as a clerk, but be made a Francis Galton Scholar. She is quite capable of doing original work.” Besides their contributions to statistics, they could become leaders in local social work. “It is most desirable that people trained in the Eugenics Laboratory should pass into work in public or municipal service of some type, as in dealing with mental defectives or invalid children, etc. We shall thus develop into a training school for practical eugenic work.”26 The most prominent of these women, F. N. (Florence Nightingale) David, named for the famous health reformer, went on to a storied career in statistics, including as professor in California.
继承与社会政策
Inheritance and Social Policy
所有这些数据工作证明了什么？情报是继承而来的，而英国正在输掉情报游戏：“作为一个国家，我们正在停止像过去那样培育情报五十到一百年前，智力较好的种群不再以与过去相同的速度繁殖。能力较差、精力较弱的种群比智力较好的种群更能生育。”这对社会改革产生了重大影响，因为问题不在于学校，而在于种群。“任何更广泛或更彻底的教育计划都无法将智力遗传弱点提升到遗传优势的水平。”唯一的“补救措施”是“改变社会中好种群和坏种群的相对生育能力。” 27
What did all this labor with data prove? Intelligence was inherited, and Britain was losing the game of intelligence: “we are ceasing as a nation to breed intelligence as we did fifty to a hundred years ago. The mentally better stock in the nation is not reproducing itself at the same rate as it did of old; the less able, and the less energetic, are more fertile than the better stocks.” This had major implications for social reform, as the problem wasn’t schools but the breeding stock. “No scheme of wider or more thorough education will bring up in the scale of intelligence hereditary weakness to the level of hereditary strength.” The only “remedy” is “to alter the relative fertility of the good and the bad stocks in the community.”27
对于皮尔逊来说，统计学是新优生社会主义的核心，而优生社会主义对于工业化和种族冲突的现代化必不可少。然而，如果目标是为优等民族制定优生计划，那么它们并不是将种族主义和阶级主义信仰体系简单地强加于数据。对头骨的研究导致皮尔逊和他的合作者艾丽斯·李否认颅骨大小与智力之间存在任何可靠的相关性，并否认头骨表现出女性天生的低智商。28
For Pearson, statistics was to be central for a new eugenic socialism necessary for a modernity both industrial and a conflict of races. If the goals were eugenic planning for a superior race, however, they were not trivial impositions of a racist and classist belief system onto data. Investigating skulls led Pearson and his collaborator Alice Lee to deny any reliable correlation between cranial size and intelligence, and to deny that skulls demonstrate the innate lower intelligence attributed to women.28
优生统计学告诉我们严酷的事实：“我们未能认识到，在现代国家斗争中，构成国家支柱的心理特征并非由家庭、学校和大学培养出来，而是骨子里培养出来的；而在过去四十年里，国家的知识分子阶层……已经不再为我们提供适当数量的人才，来开展我们帝国不断发展的工作，在日益激烈的国家斗争中冲锋陷阵。” 29当时紧迫的政治问题需要卓越的优生知识：
Eugenic statistics told tough truths: “we have failed to realize that the psychical characters, which are, in the modern struggle of nations, the backbone of a state, are not manufactured by home and school and college; they are bred in the bone; and for the last forty years the intellectual classes of the nation. . . . have ceased to give us in due proportion the men we want to carry on the ever-growing work of our empire, to battle in the fore-rank of the ever intensified struggle of nations.”29 The pressing political issues of the day required superior eugenical knowledge:
移民问题对于合理地教授国家优生学至关重要。如果在任何时候，法律都不允许人类优越于其他种族，那么努力为人类优越于其他种族制定法律又有什么意义呢？会被劣等种族的移民大批涌入所淹没，而这些移民急于从改良人类的更高文明中获益？对于优生主义者来说，允许无差别的移民是、也必然是一切真正进步的破坏性因素。30
The whole problem of immigration is fundamental for the rational teaching of national eugenics. What purpose would there be in endeavouring to legislate for a superior breed of men, if at any moment it could be swamped by the influx of immigrants of an inferior race, hastening to profit by the higher civilisation of an improved humanity? To the eugenist permission for indiscriminate immigration is and must be destructive of all true progress.30
与高尔顿一样，皮尔逊认为“民族斗争”太重要了，不能依赖于错误的优生科学：这场斗争需要更好的科学。
Like Galton, Pearson argued that the “struggle of nations” was simply too important to rest on false eugenical science: that struggle required better science.
大数据的高级科学
A Superior Science of Big Data
社会科学和生物科学需要以数学和数据生成为基础进行重塑：“老一辈生物学家松散的定性或描述性推理必须让位于精确的数理统计逻辑。训练有素的生物学家可以发现和汇总事实，就像今天的物理学家一样，但需要训练有素的数学家来推理。未来的伟大生物学家将像今天的伟大物理学家一样，是一位训练有素、有教养的数学家。” 31不用说，许多当代生物学家不同意。皮尔逊推崇大规模的数据收集和分析，而不是小规模的实验室和实验工作。
The social and biological sciences needed a remaking based in mathematics and in the production of data: “the loose qualitative or descriptive reasoning of the older biologists must give way to an accurate mathematico-statistical logic. The trained biologist may discover and tabulate facts, much as the physicist does today, but it will need the trained mathematician to reason upon them. The great biologist of the future will be like the great physicist of to-day, a mathematician trained and bred.”31 Many contemporary biologists disagreed, needless to say. Pearson extolled large-scale data collection and analysis rather than small-scale laboratory and experimental work.
生物学中的情况在政治中更是如此。皮尔逊不耐烦地指出，人们很容易对社会问题发表意见：“每个政客、每个政坛演说家，即使对天文物理学或细胞学问题发表意见，也会对出现的每个社会问题给出明确的答案。”但社会问题比天文问题困难得多。“社会问题需要科学答案。每个社会问题都属于一个最难的问题类别——它是生命问题，不是物理问题，它是生物问题，它是医学问题，它是统计的。与任何学术上的物理或生物问题相比，解决它需要的研究不亚于解决它，而是要多得多。” 32皮尔逊的实验室提供了按照这些新的科学思路组织政治和社会秩序的模型。33
What was true in biology was even more true in politics. Pearson noted with irritation the ease with which people opine on social questions: “every politician, every platform orator, who would hesitate to express even his opinion regarding a question in astronomical physics or cytology is ready with a decisive answer to each social problem that arises.” But social problems were far harder than astronomical ones. “Social problems needed scientific answers. Every social problem belongs to a class embracing the hardest of all problems—it is vital not physical, it is biological, it is medical, it is statistical. It needs not less but far more investigation for its solution than any academic physical or biological problem.”32 Pearson’s laboratories offered models for organizing political and social order along these new scientific lines.33
相关性，而非因果关系
Correlation, Not Causation
我们一直被教导，相关性并不等于因果关系。而对于高尔顿的思想继承者来说，这就是它如此令人兴奋的原因。卡尔·皮尔逊解释说，他意识到有一个“比因果关系更广泛的类别，即相关性，而因果关系只是它的极限。”现在更多的科学可以数学化：“这种新的相关性概念将心理学、人类学、医学和社会学在很大程度上带入了数学处理领域。” 34在查看没有明确因果关系的数据集时，相关性尤其具有吸引力。在研究进化时，相关性有助于理解进化过程，而无需提供其原因的知识。皮尔逊认为生育率与智力低下、道德低下密切相关。例如，相关性对于理解一个国家应该遵循的生育政策至关重要，如果它不想衰落的话。他在晚年宣称，相关性“不仅极大地拓宽了定量方法和数学方法的应用领域，而且同时改变了我们的科学哲学，甚至生命哲学本身。” 35
Correlation, we are ever taught, doesn’t equal causation. And to Galton’s intellectual heir, that’s why it was so exciting. Karl Pearson explained that he realized there was a “category broader than causation, namely correlation, of which causation was only the limit.” Now more sciences could be made mathematical: “this new conception of correlation brought psychology, anthropology, medicine and sociology in large parts [sic] into the fields of mathematical treatment.”34 Correlations were particularly attractive in looking at sets of data with no clear causal relation. In studying evolution, correlation helped understand the processes of evolution without providing knowledge of its causes. Pearson believed fertility was strongly correlated with lower intelligence, lower morals. Correlation was essential, for example, to understand the reproductive policy a nation should follow if it was not to decline. Correlation, he proclaimed late in life, “has not only enormously widened the field to which quantitative and therefore mathematical methods can be applied, but it has at the same time modified our philosophy of science and even of life itself.”35
二十世纪的大多数统计学都集中于因果关系，我们将在后面的章节中看到这一点。但我们当前的数据革命在很大程度上涉及相关性的重新出现，相关性是商业、间谍活动和科学中最重要的工具。无论是寻找相关性还是声称自己是社会世界的专家，皮尔逊精神都渗透在数据科学中。
Most of statistics in the twentieth century focused centrally upon causation, as we will see in the chapters to follow. But much of our current data revolution involves the reemergence of correlation as the most important tool in commerce, spycraft, and science. Whether in finding correlations or claiming expertise about the social world, a Pearsonian spirit pervades the data sciences.
新的数据驱动种族主义
New Data-Driven Racisms
从我们的角度来看，所有这些人物似乎都是落后的种族主义者和阶级主义者。他们确实是。但他们不是顽固的传统主义者或保守主义者。相反，他们的科学是他们进步主义的核心，是他们研究社会差异和促进民族团结的核心，他们认为民族团结的基础是他们那个时代最好的知识。这些新科学将颠覆社会秩序的概念基础——即使它们最终不会对这个秩序产生太大影响。本书的主题是展示激进的技术颠覆如何经常加剧现有的不平等。
From our point of view, all these figures appear largely backward racists and classists. And they were. Yet they were not hidebound traditionalists or conservatives. To the contrary: their science was central to part of their progressivism, to how they proposed to study social difference, and to foster a national unity that they believed to be undergirded by the best knowledge of their day. These new sciences would disrupt the conceptual foundations for the social order—even if they didn’t ultimately change this order very much. Showing how radical technical disruptions often serve to reinforce existing inequalities will be a theme throughout this book.
优生学家们发现，他们所推崇的政策很少像他们所希望的那样迅速得到实施，因此一些历史学家否认了这一运动的重要性。历史学家罗伯特·奈解释说：“优生学在英国的长期重要性在于，它将狭隘的阶级观念转变为一系列声称代表整个社会利益的生物医学概念，并成为几代受过教育的英国人无法抗拒的观点。” 36优生学思想成为许多受过教育阶层的默认思想；在英国，阶级问题占主导地位；在美国，种族问题占据突出地位。优生学思想也影响了纳粹德国的政策，导致了种族灭绝。
Eugenicists saw few of their favored policies adopted as quickly as they would have liked, so some historians have dismissed the significance of the movement. Historian Robert Nye explains, “the long-term importance of a eugenics discourse in England was the way it transformed a narrow class outlook into a matrix of biomedical concepts claiming to represent the interests of the whole society, and which became an irresistible perspective for generations of educated Britons.”36 Eugenical ideas became default ideas for many in the educated classes; in Britain concerns with class predominated; in the United States race figured prominently. Eugenical ideas also shaped policy in Nazi Germany, with genocidal results.
生物统计学、种族和现代社会问题
Biometry, Race, and the Problems of Modern Society
“如果现代文明与其他文明的区别在于其科学基础，”布拉金德拉纳特·西尔解释道，“那么现代文明所提出的问题必须通过科学的方法来解决。”在开幕致辞中，1911 年，第一届世界种族大会在 WEB Du Bois 的出席下召开，Seal 认为，解决现代世界种族问题的迫切需要新的人文科学——不是亚里士多德或马基雅维利的旧人文主义或哲学方法，而是新的生物统计科学。“对种族和民族的组成要素和构成、起源和发展以及控制这些因素的权力进行科学研究，将为在健全进步的基础上解决种族间索赔和冲突指明道路”，在分裂的美国、动荡的大英帝国和世界其他地区都是如此。37
“If modern civilisation is distinguished from all other civilisations by its scientific basis,” Brajendranath Seal explained, “the problems that this civilisation presents must be solved by the methods of Science.” In this opening address to the 1911 First Universal Races Congress, with W. E. B. Du Bois in attendance, Seal argued that the solutions to the pressing problems of race in the modern world required new sciences of humanity—not the old humanistic or philosophical methods of an Aristotle or a Machiavelli, but the new biometric sciences. “A scientific study of the constituent elements and the composition of races and peoples, of their origin and development, and of the forces that govern these, will alone point the way to a settlement of inter-racial claims and conflicts on a sound progressive basis,” in the divided US, the restive British Empire, and the rest of the world.37
西尔支持优生学计划，指出“研究遗传条件和原因，研究塑造和控制人类种族兴起、发展和衰落的生物、心理和社会权力，只有这样，我们才能通过有意识的选择，智能地适应自然系统和程序，来指导和控制人类未来的进化。” 38然而，西尔不相信人类通常的种族划分，并呼吁使用生物统计学，根据数据正确地划分人类的种族。西尔深受高尔顿和皮尔逊方法的影响：我们必须“采用生物统计学方法来研究性格和变异”，不相信平均值，因为“性格的变异范围与性格本身一样重要。” 39
Embracing a eugenics program, Seal noted, the “study of genetic conditions and causes, of the biological, psychological, sociological forces at work, which have shaped and governed the rise, growth, and decadence of Races of Man, can alone enable us to guide and control the future evolution of Humanity by conscious selection in intelligent adaptation to the system and procedure of Nature.”38 And yet Seal distrusted the usual division of humans into races, and called for biometry to delineate properly the divisions of humanity, based on data. Seal was imbued with the approaches of Galton and Pearson: we must “adopt biometric methods in studying characters and variations,” distrust averages, as “the range of variations in a character is as important an index as the character itself.”39
几年后，西尔告诉普拉桑塔·马哈拉诺比斯：“你必须在印度做卡尔·皮尔逊在英国所做的工作。”在建立机构和开展生物统计调查方面，马哈拉诺比斯做到了这一点。他将生物统计计划带到了印度，发展并挑战了皮尔逊的方法，并在印度创立了数理统计学。40
A few years later Seal told Prasanta Mahalanobis, “You have to do work in India similar to that of Karl Pearson in England.” In building institutions and pursuing biometric investigations, Mahalanobis did so. He brought the biometrical program to India, developed and challenged Pearson’s methods, and founded mathematical statistics in India.40
马哈拉诺比斯致力于获取生物特征数据并对其进行更为严格的研究，最终成为了一位极具争议的殖民者。随着印度获得独立，他把英国人创造的数据转化为强有力的民族主义知识形式。随着时间的推移，他让殖民数据服务于新的后殖民印度国家。41
Committed both to the acquisition of biometrical data and the development of ever more rigorous investigation of it, Mahalanobis came ultimately to turn highly problematic coloniai data produced by the English into potent forms of nationalist knowledge as India secured its independence. In time, he made colonial data serve the new postcolonial Indian state.41
马哈拉诺比斯追随西尔的愿望，寻求辨别不同人群的种族和种姓混合的技术。如今，他最出名的是他首先开发的一种统计学距离测量方法，即“种姓差异”，作为皮尔逊对种族差异进行科学研究的替代方法。与当时的许多种族理论家不同，西尔和马哈拉诺比斯强调的是随着时间的推移缓慢但真实的变化。在 1925 年对英孟加拉人的研究中，马哈拉诺比斯发现了戏剧性但可以理解的变化。他声称种姓具有一些短暂的现实，但“种姓综合”正在顺利进行。“省内的混合缓慢而稳定地进行着，即使不知不觉，也发展出了一个更大的印度社会，它不仅与吠陀或古典时代的传统社会不完全相同，而且在许多方面甚至是对立的。” 42数据分析揭示了一个新印度民族的缓慢生物学创造，该民族具有真正的生物学统一性，摆脱了种姓和宗派分裂。
Following the aspirations of Seal, Mahalanobis sought techniques for discerning racial and caste mixture of various populations. Today he is best known for a measure of distance used in statistics that he first developed, “caste difference,” as an alternative to Pearson’s approach to the scientific study of racial difference. Unlike many of the racial theorists of their time, Seal and Mahalanobis stressed slow but real transformations over time. In his 1925 study of the Anglo-Bengals, Mahalanobis discerned dramatic but intelligible change. Caste had some transitory reality he claimed but “caste-synthesis” was well under way. “Intermixture within the province has gone on slowly and steadily even if imperceptibly and a larger Hindu Samaj has evolved which is not only not identical with the traditional society of Vedic or classic times but is in many respects even antagonistic.”42 The data analysis revealed the slow biological creation of a new Indian nation with real biological unity out of caste and sectarian division.
帮助他量化种姓的方法是一种强大的新工具，可用于大规模研究社会群体之间的相关性。他认为，他的新实证技术既揭示了这种缓慢的统一，也揭示了印度种姓和部落的多样性。在一项大规模数据分析中，马哈拉诺比斯和他的同事对北方邦的种姓和部落进行了数据驱动的聚类。
Aiding his approach to quantifying caste were powerful new tools for examining correlations among social groups at great scale. His new empirical techniques, he argued, both revealed this slow unification but equally the diversity of castes and tribes in India. In a massive data analysis, Mahalanobis and his collaborators undertook a data-driven clustering of castes and tribes in Uttar Pradesh.
进行这些分析不仅需要一支人工计算团队，还需要使用位于英国剑桥的“马洛克机器”。43马哈拉诺比斯和他的团队在数据科学出现之前就已经开始进行数据科学研究，他们采用的实证方法侧重于计算大型学校的相关性，并使用新的计算设备。他们发现了以下两点之间的明显区别：婆罗门、工匠和部落群体。然而，尽管马哈拉诺比斯的技术非常强大，但他也认识到这些数字差异的影响是有限的。“要取得进一步进展，必须考虑到部落和种姓的社会和文化历史，也就是已知的人种学证据。” 44未能求助于这些专业知识将困扰——事实上直到今天——太多数据驱动的科学。无论算法多么强大或数据多么丰富，如果不能将这种数据分析嵌入更广泛的知识形式（科学和人文知识）中，那么所谓的知识至少应该被视为不完整的，最坏的情况是危险的。
Producing these analyses involved not just a team of human calculators, but also the use of “Mallock’s Machine” housed in Cambridge in the United Kingdom.43 Both in their empirical approach focused on calculating correlations at huge schools and in the use of new calculating devices, Mahalanobis and his team were doing data science long before data science. They found clear distinctions between Brahmins, artisans, and tribal groups. And yet, for all the power of his techniques, Mahalanobis recognized the limits of the implications of these numerical differences. “To make further progress, it is necessary to take into consideration the social and cultural history of the tribes and castes, that is, the known ethnological evidence.”44 The failure to turn to such expert knowledge would plague—and indeed plagues to this day—too much data-driven science. No matter how powerful the algorithm or extensive the data, if one fails to embed this data analysis within broader forms of knowledge, scientific and humanistic alike, that so-called knowledge should be seen as incomplete at least, dangerous at worst.
渴望找到种族和阶级差异的原因
Yearning for Causes—of Racial and Class Difference, for Example
1911 年第一届世界种族大会上，印度知识分子布拉金德拉纳特·西尔 (Brajendranath Seal) 展望了各国的团结，美国社会学家兼代表 WEB 杜波依斯 (WEB Du Bois) 回顾了大会，并从中得出了最重要的结论。“历史证明了这些真理，”他在笔记中写道，然后引用了一位杰出演讲者的话。“如果我们发现非洲某个种族的思想与欧洲种族的思想之间存在巨大差异，那么我们必须从外部条件而不是民族素质的差异中寻找原因。”“这不是种族心态的差异，而是教育的差异，我们发现，在同一种族的不同阶级之间或其历史的不同时期，也或多或少存在这种差异。” 45种族和阶级差异不应被视为理所当然，当前智力的差异不应归咎于现有的差异。
Reviewing the 1911 First Universal Races Congress where the Indian intellectual Brajendranath Seal envisioned the coming together of nations, the American sociologist and delegate W. E. B. Du Bois drew out the most significant takeaways. “History illustrates these truths,” he wrote on his notes, before quoting a distinguished speaker. “If we find an immense difference between the mind of some race” in Africa “and that of European race, we must seek the cause not in any difference of national qualities,” but in external conditions. “It is not a difference of mentality in the race, but a difference of instruction, the same difference that we find to a greater or less extent, between the various classes of one and the same race or the different periods of its history.”45 Race and class differences must not be taken for granted, and differences in current intelligence ought not to be ascribed based on existing differences.
Data, Intelligence, and Policy
早在纳粹围绕种族科学建立国家的几十年前，一名美国保险公司的员工就声称自己掌握了可以证明“雅利安人种”天生优越性的数据。1896 年，在美国经济协会的支持下，移民到美国的德国人弗雷德里克·霍夫曼发表了一本书，书中对 19 世纪下半叶美国黑人的残酷描述。在霍夫曼看来，这些数据粉碎了约翰·斯图尔特·密尔等自由派人物自鸣得意的平等主义。密尔强调性别和种族平等。但霍夫曼认为，数据毫无疑问地证明了事实并非如此。他坚持认为，无论是在欧洲殖民地还是在美国南部，政府和企业政策都必须注意不平等的科学。
Decades before the Nazis built a state around their racial science, an American insurance employee claimed to have data demonstrating the inherent superiority of the “Aryan race.” In 1896, a German immigrant to the United States, Frederick Hoffman, published a brutal portrayal of Black Americans in the second half of the nineteenth century, under the auspices of the American Economic Association. To Hoffman’s eyes, the data annihilated the smug egalitarianism of liberal figures such as John Stuart Mill. Mill stressed the equality of the sexes as well as races. Data, according to Hoffman, proved indubitably otherwise. And government and corporate policies must take heed of the science of inequality, he insisted, whether in European colonies or in the American South.
霍夫曼写道：“只有通过彻底分析构成这个国家有色人种历史的所有数据，才能理解所谓‘黑人问题’的真正本质，并安全地将过去的经验应用于解决这个国家现在面临的困难。” 1三百页之后，作者转向数据所显示的内容，“我们发现，在世界各地、所有时代和所有民族中都观察到的事实的解释不是生活条件，而是种族和遗传，即一个种族比另一个种族优越，雅利安人种总体上也比另一个种族优越。” 2霍夫曼并没有将自己的研究局限于美国的黑人。他一再坚持认为，数据确凿地表明，世界各地的殖民地人民，如美国的黑人，死亡率更高，生活水平更低，这不是由于环境或社会条件，而是霍夫曼说，他们天生的劣势。
“Only by means of a thorough analysis of all the data that make up the history of the colored race in this country,” wrote Hoffman, “can the true nature of the so-called ‘negro problem’ be understood and the results of past experience be applied safely to the solution of the difficulties that now confront this country.”1 Three hundred pages later, the author turned to what the data showed, “It is not in the conditions of life, but in race and heredity that we find the explanation of the fact to be observed in all parts of the globe, in all times and among all peoples, namely, the superiority of one race over another, and of the Aryan race overall.”2 Hoffman didn’t limit himself to Blacks in the United States. The data, he maintained time and again, showed conclusively that colonized people worldwide, like Blacks in the US, had higher mortality and lower standards of living, not due to any environments or societal conditions, but, said Hoffman, to their innate inferiority.
十九世纪下半叶，过去的旧种族主义寻求以人类学、社会学和统计学等新领域为基础的科学来获得新的合法性。这些种族科学为美国制定剥夺黑人权利的一系列法律和做法提供了掩护，即所谓的“吉姆·克劳法”。当今所谓的“种族现实主义者”延续了这种传统，用科学的外衣来粉饰偏见和系统性不平等。
In the second half of the nineteenth century, the old racisms of the past sought new legitimation in sciences grounded in the new fields of anthropology, sociology—and statistics. And these racial sciences provided cover for the creation of the broad array of laws and practices disenfranchising Blacks in the US known as Jim Crow. So-called “race realists” of today continue this heritage of dressing up prejudice and systemic inequality in scientific garb.
霍夫曼是一名雇佣兵。3保诚保险公司雇佣他来抵御反歧视法，该法禁止保险公司向黑人客户收取更高的费用——履行美国宪法第十四修正案中平等保护的承诺。他备受赞誉的工作旨在向他的雇主表明黑人根本无法投保。他声称，这些数据表明他们在达尔文式的生存斗争中失败了。霍夫曼利用广泛的数据来源，将说明黑人和白人不同死亡率的任务变成了种族等级制度的科学陈述，并警告种族混合的危险性。
Hoffman was a hired gun.3 The Prudential Insurance Company had employed him to fend off anti-discrimination laws forbidding insurers to charge Black clients more— honoring the promise of equal protection of the Fourteenth Amendment to the US Constitution. His much-celebrated work purported to show to his employers that Blacks were simply uninsurable. The data demonstrated, he claimed, that they were failing in the Darwinian struggle for existence. Drawing upon a wide array of data sources, Hoffman turned the task of illustrating the different mortality rates of Blacks and whites into a supposedly scientific statement of racial hierarchy, with warnings against the dangers of racial mixing added for good measure.
评论家，尤其是有色人种学者，驳斥了霍夫曼的推理。在对这部作品的毁灭性评论中，社会学家——后来成为 NAACP 的联合创始人——WEB Du Bois 驳斥了霍夫曼对数据的选择，强调了数据在得出一般结论方面的局限性，最重要的是，他指出了关于种族的许多说法适用于所有种族的工人阶级和新移民。这些数据非但没有证明黑人和白人之间的本质差异，反而提供了他们之间社会经济差异的指标。杜波依斯指出，作者“绝没有避免统计方法的许多谬误。这种方法毕竟只是将逻辑应用于计数，无论计数多少次都不能证明背离正确推理的严格规则是合理的。”至于“种族特征”和“生活条件”，杜波依斯指出，“他似乎有责任……证明这些种族特征在被搁置至少一个世纪之后，在 1880 年至 1890 年的十年间首次采取了果断行动。” 4
Critics, notably scholars of color, demolished Hoffman’s reasoning. In a devastating review of this work, the sociologist—and later cofounder of the NAACP—W. E. B. Du Bois tore apart Hoffman’s choice of data, stressed the limitations of the data for drawing general conclusions, and, above all, showed how many claims made about race applied to working classes and recent immigrants of all races. Far from proving the essential differences between Blacks and whites, the data provided an index of the socioeconomic differences between them. Du Bois noted that the author “has by no means avoided many fallacies of the statistical method. This method is after all nothing but the application of logic to counting, and no amount of counting will justify a departure from the severe rules of correct reasoning.” As to “race traits” and the “conditions of life,” Du Bois noted, “it would seem incumbent on him . . . to prove these race traits after being held in abeyance for at least a century, first took decisive action in the decade 1880 to 1890.”4
然而，尽管杜波依斯否认黑人整体上处于劣势，但他仍接受优生学的观点，认为所有种族都有其自然的“退化” 。5
And yet, even as he denied the inferiority of Blacks as a whole, Du Bois embraced eugenicist views that all races had their share of natural “degenerates.”5
杜波依斯的统计分析是正确的，但他的分析在很大程度上被忽视了；霍夫曼确实很荒谬，但强大的利益集团有强烈的动机相信其他观点。审计和分析算法决策的方法具有很强的启发性，无论是在 1900 年还是 2022 年，但如果没有权力或宣传的配置来使它们强大，它们往往没有效果。像杜波依斯那样，仅仅正确是不够的。
DU BOIS’S STATISTICAL analysis was correct, but his analysis was largely disregarded; Hoffman was in truth ridiculous, but powerful interests had strong incentives to believe otherwise. Methods of auditing and analyzing algorithmic decision making are powerfully illuminating, whether in 1900 or 2022, but are too often without effects without some configurations of power or of publicity to make them powerful. It isn’t enough to be right, as Du Bois was.
霍夫曼的统计分析，用历史学家乔治·弗雷德里克森的话来说，成为“十九世纪末出现的最具影响力的种族问题讨论”。正如他的雇主所希望的那样，霍夫曼的工作为拒绝在二十世纪初为非裔美国人提供保险提供了理由。世纪；这部著作和其他类似的著作为整个歧视和剥夺公民权的机器的诞生披上了一层科学的外衣。6
Hoffman’s statistical analysis became, as in the words of historian George Frederickson, “the most influential discussion of the race question to appear in the late nineteenth century.” As his employer wanted, Hoffman’s work justified refusing to insure African Americans early in the twentieth century; the work, and others like it, gave a scientific veneer to the creation of the entire apparatus of discrimination and disenfranchisement.6
霍夫曼的记录非但没有消除社会经济不平等，反而强化了这种不平等。断然拒绝为整个阶层的人提供人寿保险，产生了系统性影响，使贫困代代相传。霍夫曼并没有消除不平等的根源，而是将不平等本质化，视为自然现象。新的统计方法从根本上改变了人们对黑人身份的理解。历史学家哈利勒·纪伯伦·穆罕默德解释说，在二十世纪初，“黑人身份通过犯罪统计数据被重新塑造”。他认为，通过“种族定罪”，黑人身份成为“与白人身份相对立的更稳定的种族类别”——尤其是当以前被边缘化的移民群体——意大利人和波兰人——失去了他们可怕的名声时。7
And far from dismantling socioeconomic inequality, Hoffman’s documentation helped to reinforce it. Categorically denying life insurance to any entire class of people had systemic effects in deepening impoverishment from generation to generation. Rather than dismantling its causes, the inequality was essentialized, treated as natural. New statistical approaches radically altered how Blackness was conceived. The historian Khalil Gibran Muhammad explains, in the early twentieth century, “Blackness was refashioned through crime statistics.” Through “racial criminalization,” he argues, Blackness became “a more stable racial category in opposition to whiteness”—especially as previously marginalized immigrant groups—the Italians and the Poles—lost their fearsome reputation.7
统计数据不仅仅代表世界。它改变了我们对世界进行分类和观察的方式。它改变了我们对他人和自己的分类方式。它改变了世界。正如我们将看到的，当代数据科学正在以超快的速度做到这一点。
Statistics doesn’t simply represent the world. It transforms how we categorize and view the world. It transforms how we categorize others and ourselves. It changes the world. And, as we’ll see, contemporary data science does this—at hyperspeed.
霍夫曼试图用生物学来解释不平等——将其性质自然化——使其成为现实。他并不是一个伟大的统计学家。就在 20 世纪初，更优秀的统计学家提供了理解和证明人类差异的新方法。他们也容易犯杜波依斯在霍夫曼身上发现的错误。
Hoffman tried to make biology explain inequality—to naturalize its quality—to make it a thing. He was no great statistician. Right around the turn of the twentieth century, better statisticians provided new ways of understanding— and justifying—human difference. They, too, were prone to the errors Du Bois had espied in Hoffman.
尤尔：论贫困的原因（呃……相关性）
Yule: On the Causes (er . . . Correlations) of Poverty
尽管霍夫曼影响力巨大，但他在统计上却缺乏说服力。在为政策辩护时，他并没有借鉴高尔顿和皮尔逊的新工具可以预测和治疗疾病。一位名叫乌德尼·尤尔的优秀数学家，曾是卡尔·皮尔逊的雇员和同事，他将这些新工具运用到了人类差异研究之外，并将其应用于当时的重要社会问题。为了应对当时的社会焦虑，尤尔将最新的技术发展——多变量回归——转向贫困起伏的原因。
Despite vast influence, Hoffman was statistically underpowered. In making arguments for policy, he did not draw upon the new powerful tools of Galton and Pearson to predict and to prescribe. A fine mathematician and sometime employee and colleague of Karl Pearson named Udny Yule took those new tools beyond the study of human difference and applied them to vital social issues of the day. In keeping with societal anxieties of his era, Yule turned the latest technological development, regression with multiple variables, to the causes of the ebb and flow of poverty.
贫困的根源是什么？十九世纪末，英国曾展开过一场激烈的辩论，讨论哪些政策加剧或阻止了贫困。直接支持是否加剧了贫困？1834 年，议会颁布了《济贫法》，强迫任何被认为有能力的人在条件恶劣的济贫院工作，以此阻止人们“想”贫困。禁止向所有身体健全的成年人及其家人提供“外部救济”，即直接提供生存资金，而对在济贫院工作的人则提供“内部救济”。但这些救济形式如何影响贫困？“严厉的爱”方法是否像当时和现在的支持者经常争论的那样，减少了贫困？是否有数据和科学来支持长期以来主要是道德的论点？
What drives poverty? A lively debate in Britain toward the end of the nineteenth century concerned what policies increased or discouraged poverty. Did direct support encourage more poverty? In 1834, Parliament had enacted Poor Laws, to discourage the population from “wanting” to be poor, by forcing anyone deemed capable to work in workhouses with deliberately harsh conditions. “Out-relief,” the direct granting of funds for survival, was to be forbidden to all able-bodied adults and their families, in favor of “in-relief” for those working in such workhouses. But how did these forms of relief affect poverty? Did the tough-love approach curtail poverty, as proponents then as now often argue? Was there data and science to back up what had long been a largely moral argument?
19 世纪末，英国的统计学家试图用数据来回答这些问题。我们不应忽视，这在当时是多么奇怪。今天，我们希望政策制定者能够以数据为导向——或者至少假装如此。尽管今天一些批评者对疫苗和全球气候变化科学共识所依赖的数据提出质疑，但总的来说，我们希望将民主决策过程中的大量技术方面移交给拥有数据和分析手段的专家。我们共同赋予他们这种权力。
In the late nineteenth century, statisticians in England sought to use data to answer these questions. We should not lose sight of how odd this remained at the time. Today we expect policymakers to be data-driven—or at least to pretend to be. Although some critics today contest the data undergirding the scientific consensus about vaccines and global climate change, generally, we expect to turn over substantial technical facets of our democratic decision-making processes to experts armed with data and means for analyzing it. We collectively give them this power.
在政策问题上听从科学专家的意见并不是一个明显的举动。这种尊重使政策问题转化为科学问题，即被认为不受党派仇恨或哲学争论影响的问题。用历史学家阿兰·德罗西埃的话来说，“政治问题”被转化为“一种可以仲裁争议的测量工具”。8正如我们上面所看到的，霍夫曼试图为种族歧视及其消除提供科学依据。
Deferring to scientific expertise in questions of policy was not an obvious move. Such deference turns a policy question into a scientific one, one thought to be outside partisan rancor or philosophical debate. In the words of the historian Alain Desrosières, “a political problem” was translated “into an instrument of measurement that allowed arbitration of a controversy.”8 As we saw above, Hoffman sought to provide a scientific grounding for racial discrimination—and elimination.
19 世纪 90 年代，英国贫困问题辩论的双方都开始借鉴统计数据。查尔斯·布斯等改革者主张用科学的方法解决重大政策问题，这种方法不受传统政治分歧和道德观点的影响。他写道：“科学必须重新制定生命法则”。科学的方法而非宗教的方法将“引领我们直到找到政府问题的真正解决方案”。9布斯首先明确提出了贫困线的概念，将能够最低限度照顾家庭的人和无力照顾家庭的人区分开来。1894 年，布斯出版了《英格兰和威尔士的老年贫困人口》，这是一部划时代的社会描述著作，充满了数据和表格。布斯的方法依赖于实实在在的东西。他的愿景是调查整个伦敦，但他是通过团队帮助收集的精细“本地知识”来实现的。
In the 1890s, both sides in the English poverty debate began drawing on statistics. Reformers such as Charles Booth saw themselves as advocating a scientific approach to major questions of policy, an approach untainted by traditional political divisions and moral views. “Science must lay down afresh the laws of life,” he wrote. A scientific, not religious, approach will “lead us on till we find the true solution of the problem of government.”9 Booth first articulated a vision of a poverty line, distinguishing those who could minimally care for a family and those unable to do so. In 1894, Booth released The Aged Poor in England and Wales, an epochal work of social description, chock-full of data and tables. Booth’s approach depended on shoe leather. His vision aimed to survey all of London, but to do so through finely detailed “local knowledge” collected with the help of a team.
根据这些数据，他在书的结尾提出了一系列具有政治意义的主张。特别是，他否认了“外援”过于慷慨会导致贫困加剧的严厉信念：“外援的比例与贫困人口的总比例没有一般关系。” 10布斯的统计方法很快就遭到了尤德尼·尤尔的持续攻击。
Based on the data, he ended the book with a series of politically significant claims. In particular, he denied the tough-love belief that being too generous with “out-relief” went along with higher poverty: “The proportion of relief given out of doors bears no general relation to the total percentage of pauperism.”10 Booth’s statistical procedures soon came under a sustained attack by Udny Yule.
尤尔将高尔顿和皮尔逊为生物学发展的技术应用于政策问题。他将回归分析从研究遗传的工具转变为研究因果关系时将数据拟合成线的工具。11每一个心理学家、每一个政治学家、每一个进行回归分析的经济学家都在遵循尤尔的思路，他将回归分析应用于政策问题。数据分析新技术可以对社会做出与政策相关的主张。新专家需要的不仅仅是数据：他们需要强大的分析技术来表示、预测和开药方。
Yule applied the techniques Galton and Pearson had developed for biology to questions of policy. He transformed regression from a tool for studying inheritance to one for fitting lines to data in the investigation of causality.11 Every psychologist, every political scientist, every economist running a regression is working in the vein of Yule, who applied the new technologies of data analysis to make policy-relevant claims about the social world. The new experts needed more than data: they needed powerful analytical technologies, to represent, to predict—and to prescribe.
1899 年，尤尔出版了《英国贫困变化原因调查》，探讨了公共援助与贫困之间的关系。尤尔的答案是什么？与布斯的答案相反。尤尔声称，数据显示，财政援助导致贫困加剧。尤尔试图揭示贫困变化的原因。这种方法使人们不再将回归解释为预测，而是将其解释为处方；将其解释为原因知识，我们可以制定政策处方。
In 1899 Yule published “An Investigation into the Causes of Changes in Pauperism in England,” where he explored the relation between public assistance and poverty. Yule’s answer? The opposite of Booth’s. Yule claimed the data showed that financial assistance causes poverty to increase. Yule sought to reveal the causes of changes in poverty. This approach allowed one to move from interpreting a regression as a prediction to as a prescription; interpreted as knowledge of cause, we could set out policy prescriptions.
但是你如何找出某件事的起因呢？在尤尔的导师卡尔·皮尔逊看来，因果关系是过时的、过时的。皮尔逊认为，了解原因是不可能的。相反，皮尔逊赞扬了相关性的权力，以取代我们对因果知识的渴望：
But how do you figure out what causes something? Causation was old-school, dead in the eyes of Yule’s mentor Karl Pearson. Pearson believed knowledge of causes was impossible. Pearson instead celebrated the power of correlation to replace our yearning for causal knowledge:
正是这种涵盖了从绝对独立到完全依赖的所有关系的两件事之间的关联概念，才是我们必须用来取代因果关系这一旧观念的更广泛范畴。宇宙中的一切都只发生一次，没有重复的相同性。个别现象只能进行分类，我们的问题在于，一组或一类相似但不完全相同的东西（我们称之为“原因”）将在多大程度上伴随着或跟随另一组或一类相似但不完全相同的东西（我们称之为“结果”）。12
It is this conception of correlation between two occurrences embracing all relationships from absolute independence to complete dependence, which is the wider category by which we have to replace the old idea of causation. Everything in the universe occurs but once, there is no sameness of repetition. Individual phenomena can only be classified, and our problem turns on how far a group or class of like, but not absolutely same, things which we term “causes” will be accompanied or followed by another group or class of like, but not absolutely same things which we term “effects.”12
尤尔最初追随皮尔逊，但最终还是试图克服这种限制人类知识的方式。对因果关系的追求促使尤尔更加努力：创造新的数学和新技术来思考与政策相关的数据。
Initially following Pearson, Yule ultimately sought to overcome this way of limiting human knowledge. The temptation for causality led Yule to push harder: to create new math and new technologies for thinking through policy-relevant data.
尤尔认识到其中蕴含的巨大哲学危险：“研究经济现象之间的因果关系……为得出错误结论提供了许多机会。”社会和经济领域的复杂性不允许进行物理学的大规模简化。他解释说，统计学家不能“为自己做实验”，因此“他必须接受日常经验的数据，并尽可能讨论一组变化之间的关系。”与物理学家不同，他不能“将问题缩小到每次一个变化的影响。从这个意义上讲，统计问题比物理学问题复杂得多。” 13
Yule recognized the great philosophical dangers involved: “The investigation of causal relations between economic phenomena . . . offers many opportunities for fallacious conclusions.” The complexity of the social and economic realms did not allow for the massive simplifications of physics. A statistician could not, he explained, “make experiments for himself,” so “he has to accept the data of daily experience, and discuss as best he can the relations of a whole group of changes.” Unlike a physicist, he cannot “narrow down the issue to the effect of one variation at a time. The problems of statistics are in this sense far more complex than the problems of physics.”13
统计需要新的工具来调查社会的复杂性并揭示贫困加剧等社会弊病的根源。我们如何在不同时间和地点进行测量？
Statistics needed new tools for investigating the complexity of society and revealing the causes of social ills such as growing poverty. How do we measure it in different places and times?
为了开始回答这些问题，尤尔借鉴了高尔顿和皮尔逊的工具。他们的工具主要集中在生物数据上：动物世代之间的关系以及任何特定动物（众所周知，包括人类）身体部位之间的关系。尤尔特别利用回归工具，将它们转向经济现象。尤尔最终认为，当与背景知识相结合时，观察数据可用于推断因果关系和构建政策选择。
To begin answering these questions, Yule drew on Galton’s and Pearson’s tools. They had focused their tools largely on biological data: the relationships among generations of animals and among the body parts of any given animal, including, notoriously, human beings. Drawing particularly on the tools of regression, Yule turned them on to economic phenomena. Yule eventually argued that observational data could be used to infer causality and structure policy choices, when combined with background knowledge.
好的科学需要正视经济变化的复杂性——在这种情况下，“人们可能想到的导致贫困率变化的各种原因”。14根据尤尔的说法，可能的原因包括：
Good science would need to face up to the complexity of economic change—in this case, the “various causes that one may conceive to effect changes in the rate of pauperism.”14 According to Yule, possible causes included:
1. Changes in the method, or strictness, of administration of the law.
2. 经济状况的变化，例如贸易、工资、价格和就业的波动。
2. Changes in economic conditions, e.g., fluctuations in trade, wages, prices, and employment.
3. 一般社会性质的变化，例如某一地区的人口密度、拥挤程度或工业性质的变化。
3. Changes of a general social character, e.g., in density of population, overcrowding, or in the character of industry in a given district.
4. 更多的是道德品质的变化，例如，犯罪、非婚生、教育或可能由某些原因导致的死亡率的统计数据。
4. Changes more of a moral character, illustrated, for example, by the statistics of crime, illegitimacy, education, or possibly death rates from certain causes.
5.人口年龄分布的变化. 15
5. Changes in the age distribution of the population.15
尤尔说，第一类尤其令人感兴趣，因为这类变化“可能通过负责当局的直接行动相对迅速地实现”。
The first category is of particular interest, Yule said, because, then, change “may be comparatively rapidly effected by the direct action of the responsible authorities.”
但该如何调查呢？
But how to investigate?
从相关性到因果关系
From Correlation to Cause
尤尔借助皮尔逊和高尔顿的工具，发现救济金和贫困率实际上是高度相关的： “贫困率与救济金的比例呈正相关，即前者的平均值越高，后者的平均值也越高。这种方法似乎不容置疑。” 16
Drawing upon the tools of Pearson and Galton, Yule discerned that out-relief and pauperism were, in fact, strongly correlated: “the rate of total pauperism is positively correlated with the proportion of out-relief given, i.e., high average values of the former correspond to high average values of the latter. The method used seems to leave no room for doubt.”16
布斯则持相反观点，他分析了几个例子。尤尔批评布斯混淆了例子和整体：“布斯先生这样的统计学家竟然举出这么多例子，说明他犯了从具体事例中得出普遍结论的根本错误，这真是令人遗憾。” 17
Booth had argued the opposite, drawing on the analysis of several examples. Yule criticizes Booth’s confusion of examples for the whole: “it is extremely regrettable that a statist of Mr. Booth’s standing should have given so many examples of the fundamental mistake of founding general conclusions on particular instances.”17
然而，尤尔最初坚持要小心理解这一论点：他的主张“并没有说平均救济比例低是导致平均贫困率较低的原因，也没有说反之亦然”。他解释说：“明确地说，我的意思并不是简单地说，在一个联盟中，外在救济决定了贫困，在另一个联盟中，贫困决定了外在救济，因此，你不能平均地说出哪个是哪个：而是说，在同一个联盟中，外在救济和贫困是相互作用的。” 18
And yet Yule initially insisted on taking care in understanding this argument: his claim “does not say either that the low mean proportion of out-relief is the cause of the lesser mean pauperism or vice versa.” He explained, “To be quite clear, I do not mean simply that out-relief determines pauperism in one union, and pauperism out-relief in another, so that you cannot say which is which on the average: but I mean that out-relief and pauperism mutually react in one and the same union.”18
尤尔发现相关性本身并不够。它如何指导政策？但如何克服相关性推理不当所固有的危险？
Yule found correlation by itself inadequate. How could it guide policy? But how to overcome the dangers inherent in reasoning poorly from correlations?
标准回归的形式如下：
A standard regression would take the form:
贫困率变化=
Change in pauperism =
A + B ×（地势起伏比例变化）
A + B × (change in proportion of out-relief)
其中 A 和 B 是常数。
Where A and B are constants.
这有什么问题呢？“贫困化的变化与救济比例的变化之间的关联，可以归因于后者对前者的直接作用，也可以归因于两者与经济和社会变化的共同关联。” 19换句话说，它们可能只是因为一个共同的原因使它们一起移动而相互关联。
What’s the problem with this? “The association of the changes of pauperism with changes in proportion of out-relief might be ascribed either to a direct action of the latter on the former, or to a common association of both with economic and social changes.”19 In other words, they might correlate just because a common cause makes them move together.
尤尔试图通过在他的回归中加入一些其他特征来驯服这条恶龙：
Yule sought to tame this dragon by building into his regression a selection of other features:
尤尔试图通过这种方法找出原因，例如表明年龄分布的变化并不是普遍原因。20历史学家斯蒂芬·斯蒂格勒写道，他“利用回归方程作为一种手段，既揭示了他所寻求的关系，又允许潜在的有影响力的因素他手头上的其他变量的变化。” 21他有条不紊地进行着，直到他相信自己已经穷尽了可能的其他隐藏原因。“除非，并且直到那时，能够证明其他一些与救济率变化密切相关的量可以解释这种观察到的关联，否则就只能将结果视为政策变化对贫困变化的直接影响。” 22
With this Yule sought to isolate the causes—to show, for example, that changes in age distribution are not the common cause.20 The historian Stephen Stigler writes that he “used the regression equation as a device both to uncover the relationship he sought and to allow for potentially influential changes in the other variables he had at hand.”21 He proceeded methodically, until he came to believe he had exhausted the possible alternative hidden causes. “Unless, and until, then, it can be shown that some other quantity whose changes are closely correlated with changes in out-relief ratio can account for this observed association, there is no alternative to considering the result as indicating a direct influence of change of policy on change of pauperism.”22
完成这项任务需要大量计算，尤尔利用了 Brunsviga 机械计算器。“如果没有这些机械辅助计算工具，”他指出，“我几乎无法完成目前的工作。” 23
Undertaking this task required substantial calculation, and Yule availed himself of a Brunsviga mechanical calculator. “Without such mechanical aids to calculation,” he noted, “I could scarcely have undertaken the present work.”23
尤尔在论文的最后总结道：“总贫困率的变化总是与救济比率的变化表现出明显的相关性，但与不同联盟的人口变化或老年人口比例的变化几乎没有相关性。”谈到政策，他指出：“无论如何，观察到的贫困率变化和救济比率变化之间的大部分相关性似乎都不可能归因于政策变化对贫困率变化的直接影响。” 24然而……尤尔从未真正克服将相关性和因果关系混为一谈的危险性问题。他之后的许多学科也没有克服。他指出：“对经济现象之间因果关系的研究为得出错误结论提供了许多机会。” 25确实，正如尤尔早期最尖锐的批评者、经济学家亚瑟·庇古所指出的那样：
Yule concluded his paper: “Changes in rates of total pauperism always exhibit marked correlation with changes in out-relief ratio, but very little correlation with changes in population or in proportion of old in the different unions.” Connecting with policy, he noted: “It seems impossible to attribute the greater part, at all events, of the observed correlation between changes in pauperism and changes in outrelief ratio to anything but a direct influence of change of policy on change of pauperism.”24 And yet . . . Yule never really surmounted the concerns around the dangers of conflating correlation and causation. And neither have many of the disciplines in his wake. “The investigation of causal relations between economic phenomena,” he noted, “offers many opportunities for fallacious conclusions.”25 True that, as Yule’s sharpest early critic, the economist Arthur Pigou, noted:
研究发现，在我国的各个工会中，已经和正在推行了各种政策，并通过统计推理提出这些政策，以事后证明一种政策比另一种政策具有更好的经济效果。26
It observed that, in the various unions of this country, various lines of policy are being and have been pursued, and proposed by statistical reasoning to demonstrate a posteriori that one line of policy has better economic effect than another.26
庇古认为，“对这种统计推理形式的基本反对意见是，一些最重要的影响因素是无法量化衡量的，因此无法纳入统计机制的管辖范围，无论统计机制多么复杂。” 27
“The fundamental objection to” this form of statistical reasoning, Pigou argued, “is that some of the most important influences at work are of a kind which cannot be measured quantitatively, and cannot, therefore, be brought within the jurisdiction of statistical machinery, however elaborate.”27
不难推断出一个根本原因来解释救济金和贫困之间的相关性。而实地实际知识很容易提供一个原因。基于这一知识，尤尔 1909 年的对手庇古认为，“更好的管理”是贫困和救济金增加的基础。“鉴于这种情况，还有什么比认为救济金比例和贫困率之间观察到的相关性不是由于任何直接的因果关系，而是由于两者都是由一般管理的性质造成的更自然的呢？” 28
It’s not hard to postulate an underlying cause that would account for the correlation between out-relief and pauperism. And real knowledge on the ground easily supplied one. Based on that knowledge, Yule’s 1909 opponent Pigou suggested that “better administration” underlay increases of pauperism and of out-relief. “In view of this circumstance, what more natural than to suggest that the observed correlation between the out-relief ratio and the pauperism percentage is due, not to any direct causal connection, but to the fact that both have been caused by the character of the general administration?”28
然而，这位批评家并不是卢德分子，而是一位深思熟虑的人，他深知相关性推理的危险性。从图中可以看出两者的区别。尤尔声称前者是正确的，而他的对手则提出了另一个原因。
This critic was no Luddite, however, but a careful thinker about the dangers of reasoning from correlations. The difference can be seen in the diagram. Yule claims the first is true, while his opponent posits an alternate cause.
近一个世纪后，统计学家戴维·弗里德曼（David Freedman）在评论中赞扬了统计建模中“鞋革”所获得的知识：“统计技术很少能够取代良好的设计、相关数据以及在各种环境下根据现实测试预测。” 29
A critic writing nearly a century later, the statistician David Freedman, celebrated the knowledge gained by “shoe leather” in the face of statistical modeling: “statistical technique can seldom be an adequate substitute for good design, relevant data, and testing predictions against reality in a variety of settings.”29
尽管尤尔的方法技术精湛，但他的方法无法回答如何仅使用数学来消除相关性和因果关系的歧义。与后来的许多追随者不同，尤尔明白这一点。尽管他在论文标题中添加了“原因”一词，但一个隐藏的脚注为他提供了认识论上的逃避：“严格来说，‘由于’应该理解为‘与……相关’。” 30
For all its technical prowess, Yule’s approach had no answer as to how to use mathematics alone to disambiguate correlation and causation. Unlike many of his later followers, Yule understood this. Despite adding the word “Causes” to the title of his paper, a buried footnote provides his epistemic escape: “strictly speaking, for ‘due to’ read ‘associated with.’ ”30
虽然 Yule 的工作在关于济贫法的辩论中没有立即产生实际效果，但他的技术后来成为随着一门又一门学科寻求科学地位，多元回归分析也变得至关重要：首先是经济学，然后是心理学，然后是政治学，它们都以多元回归分析作为其专业知识必不可少的基础技术。尽管尤尔的分析并没有影响他那个时代的政策，但他的分析已经构建了我们几代人的生活现实。尽管存在逻辑和证据问题，但回归分析仍然是社会科学和政策科学中的主要工具，而且实际上经常是科学分析的必要标志。
While Yule’s work had little immediate practical effect in debates about Poor Laws, his techniques would become central as one discipline after another sought scientific status: first economics, then psychology, then political science would all come to have multiple regression as a foundational technique essential to their claims to expertise. Although Yule’s analysis didn’t affect policy in his time, analysis of his kind has structured our lived reality for generations. Despite the logical and evidential problems, regressions remain a dominant tool in the social sciences and sciences of policy, and indeed often serve as a necessary sign of something being a scientific analysis.
救济和贫困增加的其他潜在原因。我们的图表灵感来自 David Freedman 的《从关联到因果：关于统计学历史的一些评论》，《统计科学》第 14 卷，第 3 期（1999 年 8 月 1 日），第 248-89 页。
Alternate potential causes for increase in out-relief and pauperism. Our diagram, inspired by David Freedman, “From Association to Causation: Some Remarks on the History of Statistics,” Statistical Science 14, no. 3 (August 1, 1999), 248–89.
设计和代理
Design and Proxies
这场十九世纪的辩论从根本上讲是关于贫困的。像繁荣一样，贫困无法直接衡量。因此，任何希望量化贫困的人都需要选择一些东西更容易衡量的指标——一种替代指标——代表贫困。科学家必须始终做出这样的选择。虽然有必要，但这样的选择既不中立也不无问题。它们是设计选择，使知识成为可能——但也容易受到严重误解。
This nineteenth-century debate is fundamentally about poverty. Like prosperity, poverty cannot be measured directly. So, anyone wishing to quantify it needs to choose something more easily measured—a proxy—to stand in for poverty. Scientists must make such choices all the time. While necessary, such choices are neither neutral nor unproblematic. They are design choices that make knowledge possible—but also subject to dramatic misinterpretation.
在英国的辩论中，贫困的关键代表是赤贫。与贫困不同，赤贫是一个行政 类别。它不是人的品质，而是政府对人的分类方式。行政类别提供了使用固定定义对人进行分类的方法，官僚机构可以大规模管理这些定义。它们还会产生可供分析的数据集。行政类别是强有力的惯例，是分析社会所必需的，但不是自然真理，只是存在着等待被发现。法国历史学家德罗西埃解释说，像赤贫这样的对象“凭借其社会编纂而存在，通过将具有波动模式的行政过程的结果具体化”。物化是一个花哨的词，指的是将想法视为事物——字面意思是从对真实事物的抽象中创造出事物。这是一个危险的错误，但在使用统计学来思考社会、政治和商业问题时，它始终是一个危险。德罗西埃写道：“正是这种从过程到事物的转变，使得尤尔的结论难以解读。” 31
In the English debate, the key proxy for poverty was pauperism. Unlike poverty, pauperism is an administrative category. It’s not a quality of people so much as the way a government classifies them. Administrative categories provide ways of classifying people using set definitions that bureaucracies can administer at great scale. And they produce data sets open to analysis. Administrative categories are powerful conventions, necessary to analyze society, but are not truths of nature, just existing to be found. The French historian Desrosières explains that an object like pauperism “exists by virtue of its social codification, through the reification of the results of an administrative process with fluctuating modalities.” Reification is the fancy word for thinking that ideas are things—literally it means making a thing out of an abstraction about real things. It’s a dangerous mistake, yet a constant danger in using statistical work in thinking through social, political, and commercial problems. “It is this slippage from the process to the thing,” Desrosières writes, “that made Yule’s conclusion so ticklish to interpret.”31
物化涉及一种根本性的弊端，即宣称对我们有用的惯例是存在的。物化的危险在智力及其与种族关系的研究中可能产生了如此恶劣的影响。
Reification involves the fundamental vice of claiming existence for that which is a convention useful to us. And perhaps nowhere has the dangers of reification had such baleful effects as in the study of intelligence and its relationship to race.
智力测试的诞生
Birth of Intelligence Testing
统计学家 Prasanta Mahalanobis 的职责很多，其中之一就是分析学校和学院的管理。比如当时和现在，世界各地的学校管理人员都对马哈拉诺比斯感兴趣，他开始将智力测试作为教育机构录取过程的一部分。和其他智力测试者一样，他将学业成功与最初基于学业成功 (IQ) 的测试联系起来，发现它们有很好的相关性。然而，作为第一个孟加拉语智力测试的创始人，马哈拉诺比斯小心谨慎，没有像他的许多同时代人那样肆无忌惮。尽管他花了几十年的时间研究印度的种姓和部落，但他并没有大肆宣扬种姓之间的智商差异，特别是声称先天智力的差异可以解释——并证明——社会和政治地位的差异。他没有把测试变成关于智力本质的主张。他没有用它来证明自然等级制度的合理性。学者 Shivrang Setler 展示了他如何将测试视为实用工具。32
Among the many duties of the statistician Prasanta Mahalanobis was analysis of the management of schools and colleges. Like administrators across the world then and now, he turned to intelligence testing as part of the process of admission to educational institutions. Like other intelligence testers, he correlated academic success with a test originally based on academic success (IQ) and discovered they were well correlated. Creator of the first Bengali-language intelligence tests, Mahalanobis, however, was careful not to do what many of his contemporaries did with abandon. Despite his decades of work investigating castes and tribes in India, he offered no grand proclamations of IQ differences among the castes, and in particular claims that variations in innate intelligence explained—and justified—differences in social and political status. He didn’t make a test into a claim about the nature of intelligence. He didn’t use it to justify a natural hierarchy. The scholar Shivrang Setler has shown how he treated the tests as practical instruments.32
他的许多同代人远没有那么克制。他们把测试和成功衡量标准与归因联系起来。他们将代理物化为某种他们没有合理机制的东西。更糟糕的是，他们大规模地呼吁并为从反移民措施到强制绝育等优生学计划提供科学依据。
Many of his contemporaries were far less restrained. They took correlations between tests and measures of success and imputed cause. They reified proxies into something for which they had no plausible mechanisms. And worse, and at great scale, they called for and offered scientific justification for eugenics programs ranging from anti-immigration measures to forced sterilization.
1904 年，一位英国心理学家公布了一个令人惊讶的结果：即使在现代蒸汽机、电报和铁路时代，学习拉丁语和希腊语也能为领导者提供良好的资历。拉丁语和希腊语的能力与作者所说的“一般智力”最为相关。
In 1904, an English psychologist revealed a surprising result: the study of Latin and Greek offered good credentials for leaders, even in a modern age of steam engines, telegraphs, and railroads. Competency in Latin and Greek correlated most with what the author dubbed “General Intelligence.”
我们不必继续徒劳地抗议说，希腊语法的高分并不能证明一个人是否有能力指挥军队或管理省份，我们最终应该真正确定各种测量一般智力的方法的精确性，然后，我们将以同样的方式
Instead of continuing ineffectively to protest that high marks in Greek syntax are no test as to the capacity of men to command troops or to administer provinces, we shall at last actually determine the precise accuracy of the various means of measuring General Intelligence, and then we shall in an equally
以积极客观的方式确定一般智力与其他特征相比的确切相对重要性。33
positive objective manner ascertain the exact relative importance of this General Intelligence as compared with the other characteristics.33
在一篇出色的实验心理学论文中，作者查尔斯·斯皮尔曼 (Charles Spearman) 指出，一系列不同的认知能力和感知能力彼此之间有着密切的关联。然后他推断，它们都可以用一种称为g的一般智力的底层形式来理解。
In a remarkable piece of experimental psychology, the author, one Charles Spearman, illustrated that a range of different cognitive and sensory abilities were strongly correlated with each other. And then he reasoned that they could all be understood in terms of an underlying form of general intelligence, called g.
在斯皮尔曼看来，在古典学方面表现优异的人之所以聪明，并不是因为学习了古典学，而是因为在古典学方面的优秀与其他形式的优秀最为相关。在古典学方面的优秀最能说明一个人具有很高的先天智力，即g。
In Spearman’s account, someone who excelled in classics was intelligent, not due to the study of classics, but because excellence in the study of classics was the most correlated with other forms of excellence. Excellence in classics best indicated a high level of innate intelligence, or g.
在创建g时，斯皮尔曼将大量测量能力简化为一个可以排序的数值。这项工作基于新颖的数学：他展示了如何采用许多变量并识别其变化和相关性背后的潜在或“潜在”因素，这一过程称为因子分析，是现代统计学的核心。如果有人对人类排名感兴趣，这种技术非常有效，因为它需要对每个人的许多方面进行丰富的描述，并将其简化为由单个潜在因素衡量的东西。这些技术在我们的算法世界中至关重要。
In creating g, Spearman reduced a large number of measured aptitudes to a single value that could be ranked. This work rested on novel mathematics: he showed how to take many variables and discern the underlying or “latent” factors that underlie their changes and correlation, a process known as factor analysis, and central to modern statistics. This technique was potent if one was interested in ranking human beings, since it took a rich portrayal of many facets of each person and reduced them to something measured by a single underlying factor. Such techniques are central in our algorithmic world.
斯皮尔曼和许多跟随他的人又向前迈出了关键的一步。他们将这一潜在因素——能力之间相关性的抽象——变成了一种事物。他们将其“具体化”：这种相关性被转化为人类拥有的真正事物，即一般智力。而且，他认为，这种智力大体上是遗传的。
Spearman, and many who followed him, went one crucial step further. They turned this latent factor—an abstraction from a correlation among abilities—into a thing. They “reified” it: this correlation was transmogrified into a real thing humans possess, a general intelligence. And this intelligence was, by and large, he believed, inherited.
高尔顿在他的优生学计划中假设最优秀的人具有更高的自然能力，但他无法直接测量这种能力。*相反，他断言 声誉为衡量自然能力提供了一个很好的替代指标。显然，需要一种更高级的程序。斯皮尔曼提供了测量任何人智力的缺失技术，并且至关重要的是，将他们按最聪明（或最好）到最不聪明的等级顺序排列。正如一位历史学家所解释的那样，“个体差异科学是由弗朗西斯·高尔顿发明的，由卡尔·皮尔逊系统化，并由查尔斯·斯皮尔曼应用于心理学。” 34
In his eugenics program, Galton had assumed that the best people had higher natural ability, but he had no way of measuring that ability directly.* Instead, he asserted that reputation provided a good proxy for gauging natural ability. A superior procedure was clearly needed. Spearman provided the missing techniques for measuring the intelligence of anyone and, crucially, for placing them into a hierarchical order of the most intelligent (or best) to the least. As one historian explains, “The science of individual differences was invented by Francis Galton, systematised by Karl Pearson, and applied to psychology by Charles Spearman.”34
凭借这些识别智力的技术，斯皮尔曼终于让心理学跃升为一门真正的科学。1923 年，他解释说：“我们必须大胆地希望，心理学长期以来缺失的真正科学基础终于得到了补充，这样它从此就能与其他坚实的科学，甚至物理学本身一起占据应有的地位。” 35
And with these techniques for identifying intelligence, Spearman had finally allowed psychology to make the jump to a true science. In 1923, he explained, “we must venture to hope that the so long missing genuinely scientific foundation for psychology has at last been supplied, so that it can henceforward take its due place along with the other solidly founded sciences, even physics itself.”35
在讨论臭名昭著的移民限制法案——1924 年美国移民法案（约翰逊-里德法案）时，斯皮尔曼解释道，
Discussing the US Immigration Act of 1924 (the Johnson-Reed Act), a notorious restriction of immigration, Spearman explained,
几乎每位研究者都强调的一般结论是，就“智力”而言，日耳曼人种平均比南欧人种具有明显优势。这一结果似乎对塑造最近非常严格的美国移民准入法产生了至关重要的实际影响。36
The general conclusion emphasized by nearly every investigator is that, as regards “intelligence,” the Germanic stock has on the average a marked advantage over the South European. And this result would seem to have had vitally important practical consequences in shaping the recent very stringent American laws as to admission of immigrants.36
尽管如此，斯皮尔曼仍保持谨慎，并注意到教育方面的社会条件的差异：
Despite this Spearman was cautious, noting the differences in social conditions around education:
大量的证据表明，不同种族之间确实存在差异，至少在遗传方面是这样。然而，这种种族差异即使真的存在，与属于同一种族的个体之间的差异相比，无疑也非常小。37
There has been found a large body of evidence that races do differ from one another, at any rate in respect of g. . . . Nevertheless such racial differences, even if truly existing, are indubitably very small as compared with those that exist between individuals belonging to one and the same race.37
那些受到斯皮尔曼启发的人在得出结论时远没有那么克制。他们如此不克制以至于他们没有让一些学术欺诈妨碍某些种族主义科学，正如斯蒂芬·杰伊·古尔德等许多其他作者早已证明的那样。高尔顿的追随者中，许多人比斯皮尔曼更狂热地追求优生学，他们热切地采用了他的技术。38
Those inspired by Spearman proved far less restrained in their conclusions. So less restrained that they didn’t let a little academic fraud get in the way of some racist science, as numerous other authors, such as Stephen Jay Gould, have long demonstrated. Galton’s followers, many far more eugenically rabid than Spearman, eagerly adopted his techniques.38
卡尔·皮尔逊担心他的优生学家同行可能会让他们对特征遗传证据的渴望妨碍更仔细的研究。他是对的。皮尔逊攻击伦敦大学学院的同事斯皮尔曼未能证明他的一般智力假设：
Karl Pearson worried that his fellow eugenicists might let their desire for proofs of the inheritance of traits get in the way of more careful inquiry. He was right. Pearson attacked Spearman, a colleague at University College London, for failing to prove his hypothesis of general intelligence:
在进行测试之前，应该通过心理共识来选择非重叠心理能力的性质，……应该使用经过训练的计算机，如果可能的话，应该开发出更适合整个主题的数学理论。然后，我们可能更有机会驳斥整个理论，或者表明值得花更多精力来开发它。目前，我们只能返回，但肯定返回未经证实的裁决。39
The nature of the non-overlapping mental abilities should be selected by psychological consensus before the tests are made, . . . trained computers should be employed, and if possible a more adequate mathematical theory of the whole subject developed. Then we might have a better chance either of dismissing the whole theory, or showing that it was worthwhile spending further energy in developing it. At present we can only return, but return definitely, a verdict of non-proven.39
讲述智力的悲惨故事会让我们走得太远。随着智商测试的普及，20 世纪的第一个十年见证了智力评估方式的重大转变。“智力”被理解了。历史学家约翰·卡森解释说，智力突然被“理解为一种差异化、可量化、单线的实体，它决定了个人或群体的整体智力。” 40人和种族可以按照一个线性尺度排序，方便的测试可以在这个尺度上识别和定位人们，使将人们分配到学校或工作成为一项简单的任务。它仍然很迷人。
To tell the sorry tale of intelligence would take us too far afield. In conjunction with the spread of IQ testing, the first decade of the twentieth century saw a dramatic shift in how “intelligence” was understood. Intelligence, the historian John Carson explains, abruptly became “understood as a differential, quantifiable, unilinear entity that determined an individual’s or group’s overall mental power.”40 People and races could be ordered along one linear scale, and convenient tests could identify and locate people on that scale, making the assignment of people to schools or jobs a straightforward task. It beguiles still.
时至今日，科学种族主义者仍然借鉴斯皮尔曼对人类进行排名的方法，但并没有他所采用的那么多细微差别。令人惊讶的是，他们仍然在这样做，这种做法永恒地重复发生，需要时刻保持统计警惕。每一代人都会得到一条钟形曲线，自信地用科学严谨的假象来掩盖当前的不平等现象。41
To this day, scientific racists have drawn on Spearman’s approach to ranking human beings and without much of what nuance he had. Remarkably they still do, in an eternal recurrence that demands ever present statistical vigilance. Each generation gets another Bell Curve that confidently dresses up current inequalities in the illusion of scientific rigor.41
从数据渴望到数据实现
From Aspiration for Data to Realization of Data
南丁格尔、高尔顿和凯特勒梦想着系统地收集有关人的数据，以指导政策制定者；皮尔逊组织了主要由女性组成的团队，负责收集和分析数据。这些改革者将人视为稳定的官僚实体的梦想从 20 世纪初开始实现。42在20 世纪上半叶，美国和欧洲及其殖民地的人口数据收集急剧扩大。我们现在习以为常的记录形式，如出生证明，以前所未有的方式将人变成了数据。而且他们这样做只是通过大量、后果严重、有争议的工作，历史学家旺吉·穆伊盖 (Wangui Muigai) 将其描述为“关于如何解释个人以及构建和记录这些资料所涉及的劳动的互动、对抗和争议”身份。” 43正如 150 年前德国“庸俗统计”批评者所认识到的那样，将事物放入数字或简单的类别中会削弱现实和多样性。对一个国家的所有居民都这样做，不仅是将他们分类，而且构建了他们对自己的理解以及他们与法律、医疗和教育当局的关系。在美国，当局仔细地根据种族和性别类别来定位每个婴儿，以确保社会秩序。穆伊盖解释说，修订后的出生证明表格“成为一种重要工具，从出生开始，它就被用来在弗吉尼亚州区分谁算白人，谁算黑人。” 44在殖民时期的印度，同样主要关注的是将人们归入行政和种姓类别。除了出生证明和人口普查，智力和性格测试自第一次世界大战以来也蓬勃发展。所有这些都是一个不平衡的过程，受到很多抵制，经常被忽视，通常不是通过复杂的数学统计，而是通过传统的官僚工作，如文档和标准化来完成。马哈拉诺比斯曾努力对孟加拉致命的殖民晚期饥荒进行统计调查，他指出，认真的数据收集需要建立“一个高效的人力组织，配备经过精心挑选和培训的工作人员。这需要时间。如果没有这样的时间，结果往往不仅无用，甚至有害。” 45历史学家桑迪普·默蒂亚 (Sandeep Mertia) 展示了马哈拉诺比斯的努力如何“在资源和人员有限的条件下，在印度幅员辽阔、语言和社会文化多样化的地理环境中，将计算工作规模化和标准化。” 46
Nightingale, Galton, and Quetelet dreamed of the systematic collection of data on people to guide policymakers; Pearson organized teams of primarily women to undertake the labor of collection and analysis. These reformers’ dreams of rendering people as stable bureaucratic entities came to be realized from the early twentieth century onward.42 In the first half of the twentieth century, the collection of data on populations expanded dramatically, in the United States and Europe and in their colonial possessions. Forms for recording that we now take for granted, such as the birth certificate, made people into data in a hitherto unprecedented way. And they did so only through tremendous, consequential, contested work, what historian Wangui Muigai describes as the “interactions, confrontations, and disputes over how individual people should be accounted for and the labor involved in constructing and documenting those identities.”43 As German critics of “vulgar statistics” had recognized 150 years before, putting things into numerical or simple categories flattens reality and diversity. And doing so for all residents of a country not just classifies them but structures their understanding of themselves and their relationships with legal, medical, and educational authorities. In the United States, authorities sought carefully to locate each infant by a precise racial and sexual category, to secure the social order. A revised birth certificate form, Muigai explains, “became a key tool for policing, from birth, who counted as white and who counted as black in Virginia.”44 In colonial India concerns with locating people in administrative and caste categories likewise predominated. Along with birth certificates and the census, intelligence and personality testing boomed from World War I. All of this was an uneven process, much resisted and often ignored, often done less through fancy mathematical statistics than via conventional bureaucratic labors of documentation and standardization. Having struggled to undertake statistical inquiry into the deadly late colonial famine in Bengal, Mahalanobis noted that serious data collection requires building “up an efficient human organisation with carefully selected and trained staff. This takes time. And unless such time is allowed the results are often not only useless, but even harmful.”45 Historian Sandeep Mertia shows how Mahalanobis’ efforts focused “on scale and standardization of computational work in conditions of limited resources and staff, . . . in India’s large and linguistically and socioculturally diverse geography.”46
从 20 世纪 30 年代开始，美国政府统计人员引入了代表性抽样技术，以便对整个国家做出新颖的断言，而无需详尽登记每个人或农场。Emmanuel Didier 认为，“代表性调查的出现伴随、推动并确认了一种“新政府干预主义”的诞生，这种主义是罗斯福新政之后中世纪福利国家的特征。47与此同时，德国等地的威权政权越来越深入地干涉公民和企业的私事，并建立了越来越全面的监控系统。48在毛泽东的领导下，中华人民共和国拒绝了抽样，认为这是资产阶级的数学诡辩，而主张真正的共产主义“定期报告和人口普查的详尽方法”。49在所有这些不同的案例中，大量的数据收集和分析工作使得对国家、人口和经济有了全新的认识，并且具有可操作性。
From the 1930s in the United States, government statisticians introduced techniques of representative sampling to make novel claims about the nation as a whole without exhaustively registering every person or farm. Emmanuel Didier argues that “the emergence of representative surveys accompanied, informed, and confirmed the birth of a new governmental interventionism” characteristic of the midcentury welfare state in the wake of the New Deal.47 Authoritarian regimes in Germany and elsewhere at the same time plumbed ever further into the private affairs of their citizens and corporations toward ever greater total systems of surveillance.48 Under Mao, the People’s Republic of China, for their part, rejected sampling as bourgeois mathematical sophistry in favor of properly communist “exhaustive methods of periodic reports and censuses.”49 In all these diverse cases, the great labor of data collection and analysis made dramatic new understandings of the state, the population, and the economy possible—and actionable.
伴随着这些大规模数据收集工作的是允许自动处理数据的基础设施的建立，这远早于数字计算机的出现，其中最著名的是 Hollerith 打卡机。历史学家亚当·图兹认为，“由于打卡机的存在，世界各地的官僚们都梦想着无所不知”：“这是第一次有可能想象整个国家的数据都记录在一个数据库中，通过机械处理设备可以立即访问。” 50一本早期的世界统计学史指出：“这样一来，人口统计数据的更精细的统计呈现方式的社会价值尚未完全被理解。”作者解释说，由于没有打卡机，“如果我们要成功应对由我们有缺陷的移民法强加给我们的种族混合而产生的问题，我们就永远不可能揭露我们必须知道的所有真相。” 51然而，无论它们如何扁平化人类生存的结构，将人类生活转化为机器处理的统计数据仍然会使分层变得可见，这既可以证明和延续这种分层，也可以对其进行质疑。
Alongside these massive data collection efforts came the making of infrastructures allowing for the automatic processing of data, long before digital computers, most notably Hollerith punched card machines. “Across the world,” historian Adam Tooze argues, “bureaucrats were inspired to dreams of omniscience” thanks to the card machines: “For the first time it became possible to conceive of an entire nation recorded in a single database instantly accessible by means of mechanical handling equipment.”50 An early history of statistics worldwide noted, “The sociological value of the minuter statistical presentation of demographic data thus brought within reach, is not yet fully understood.” Lacking the punched card machines, the authors explained, “we could never hope to lay bare all the truth we must have, if we are to cope successfully with the problems growing out of the heterogeneous commingling of races which our defective immigration laws are forcing upon us.”51 And yet, however much they flattened the texture of human existence, such rendering of human life into statistics processed on machines nevertheless could make stratification visible, both to justify and extend that stratification—or to contest it.
科学种族主义的傲慢
The Hubris of Scientific Racism
在所有逃避考虑社会和道德影响对人类心灵的影响的庸俗方式中，最庸俗的就是将行为和性格的多样性归因于固有的自然差异。
Of all vulgar modes of escaping from the consideration of the effect of social and moral influences on the human mind, the most vulgar is that of attributing the diversities of conduct and character to inherent natural differences.
—约翰·斯图尔特·密尔52
—John Stuart Mill52
统计学的大部分历史都与一个漫长而令人遗憾的故事交织在一起，即人们试图证明社会等级制度建立在人与人之间天生的差异之上，无论是性别、种族还是阶级。我们一次又一次被这种说法欺骗，这种说法一直延续到我们的基因组时代。53很少有科学主张比声称天生差异恰好反映了我们当前的社会安排更值得怀疑。历史告诉我们，这种说法需要最高级别的警惕——关于所使用的数据、如何操纵数据以及从中得出的推论。通常教训很简单：我们所知道的确切数字远没有许多提供统计数据和统计推论的人声称的那么准确，我们需要我们自己的杜波依斯来提醒我们。
Much of the history of statistics intertwines with the long, sorry tale of attempts to prove that social hierarchies rest on innate differences between people, whether differentiated by sex, race, or class. We’ve been duped time and again by such claims, which have persisted to our genomic age.53 Few scientific claims should be viewed with more suspicion than claims to innate difference that just happen to reflect our current social arrangements. History teaches that such claims demand the highest level of vigilance—about the data used, how that data is manipulated, and the inferences drawn from it. Often the lesson is simple: we know far less with certainty than many people proffering statistics and statistical inferences claim, and we need our own Du Boises to remind us.
霍华德大学数学家凯利·米勒在评论霍夫曼的种族统计处理时，猛烈批评了其糟糕的推论和错误的数据，以及未能考虑其他解释：“它不能像在其他假设下那样令人满意地解释其所安排的事实。作者没有考虑到令人沮丧的观察事实可能是由于解放和重建的剧烈动荡造成的，因此，它们只是暂时的。” 54
In his review of Hoffman’s statistical treatment of race, the Howard University mathematician Kelly Miller savaged the poor inferences and faulty data, and failure to consider alternate explanations: “It does not account for the facts arranged under it as satisfactorily as can be done under a different hypothesis. The author fails to consider that the discouraging facts of observation may be due to the violent upheaval of emancipation and reconstruction, and are, therefore, only temporary in their duration.”54
尽管数据和新技术呈爆炸式增长，但20 世纪初，没有人知道如何最好地检验这些不同的、相互竞争的假设。需要创造一种新的科学形式来回答紧迫的问题：哪种肥料促进了大麦的生长？哪种药物最有效？而这种科学首先在吉尼斯啤酒厂诞生，由历史上被称为“学生”的人创造。
Despite the explosion of data and new techniques, at the turn of the twentieth century, no one knew how best to test such different, competing hypotheses. A new form of science needed to be created, to answer pressing questions: Which fertilizer encouraged the growth of barley? Which medicine worked most effectively? And that science was first created at the Guinness Brewery, in the hands of someone known to history as “Student.”
* “我所说的天赋是指智力和性格的品质，它们促使一个人有能力并使其有资格从事能带来名声的行为。我指的不是没有热情的能力，也不是没有热情的能力，甚至也不是两者的结合，没有足够的能力来完成大量非常艰苦的工作……天赋在自然状态下，会在内在刺激的推动下，爬上通往卓越之路，并有权力到达顶峰——如果受到阻碍或挫败，它会焦躁不安，努力奋斗，直到克服障碍，再次自由地遵循其热爱劳动的本能。”弗朗西斯·高尔顿，《遗传天才：探究其规律和后果》（伦敦：麦克米伦，1869 年），第 37、38 页。
* “By natural ability, I mean those qualities of intellect and disposition, which urge and qualify a man to perform acts that lead to reputation. I do not mean capacity without zeal, nor zeal without capacity, nor even a combination of both of them, without an adequate power of doing a great deal of very laborious work . . . nature which, when left to itself, will, urged by an inherent stimulus, climb the path that leads to eminence, and has strength to reach the summit— one which, if hindered or thwarted, will fret and strive until the hindrance is overcome, and it is again free to follow its labour-loving instinct.” Francis Galton, Hereditary Genius: An Inquiry into Its Laws and Consequences (London: Macmillan, 1869), 37, 38.
1886 年，健力士股票首次公开募股，引发了巨大轰动，有抱负的投资者纷纷打开银行大门，向历史悠久的巴林银行投机。爱德华·健力士出售了其在公司 65% 的股份，获利 600 万英镑，公司估值因此达到 900 万英镑左右——以今天的美元计算，远远超过 3000 亿美元。不管这是否是愚蠢的投机行为，健力士突然间拥有了充足的资源，可以使用当时的最新技术——统计学——来改造其业务。在数据科学有望彻底改变商业实践的一百年前，健力士团队就试图创建一门工业啤酒酿造科学。和如今硅谷或深圳的富有公司一样，健力士资金充裕，聘请了才华横溢的年轻科学家和工程师，任命他们担任“酿酒师”的尊贵职位，并为他们建造了新的实验设施。当时和今天一样，新的科学技能有可能取代旧的专业知识——而且事实往往如此。在新的化学和数学工具面前，旧形式的农业知识和专业知识逐渐淡出。
So enormous was the excitement about the initial public offering of Guinness stock in 1886, that aspiring investors broke the door of their banker, the storied firm of Barings. Selling 65 percent of his stake in the company, Edward Guinness gained six million pounds, thus giving the company a valuation of some nine million pounds—well over $300 billion in today’s dollars. Speculative folly or not, Guinness suddenly was well-resourced to transform its business using the latest technology of the day: statistics. One hundred years before the data sciences promised to revolutionize business practices, the team at Guinness tried to create an industrial science of brewing. Flush with cash like the rich firms in Silicon Valley or Shenzhen today, Guinness hired talented young scientists and engineers, appointed them to the august position of “brewers,” and built them new experimental facilities. Then as today, new scientific skills threatened to render older forms of expertise obsolete—and often did so. Older forms of agricultural knowledge and expertise faded in the face of new chemical and mathematical tools.
无论是检测啤酒花、大麦还是不同肥料的效果，吉尼斯研究的科学家们都遇到了两个主要困难：他们只有少量的观察结果，而且这些观察结果在各个研究对象之间差异很大。他们自己。1他们需要某种方法来衡量哪些差异很重要，知道哪些差异是显著的。如果你愿意的话，可以称之为显著性检验。其中一人，威廉·戈塞特，更擅长数学。显著性检验就是他发明的。这一切都是为了酿造出更好、更赚钱的啤酒。
Whether examining hops, barley, or the effects of different manures, the Guinness research scientists ran into two major difficulties: they had only a small number of observations and those observations varied considerably among themselves.1 They needed some way to gauge which differences mattered, to know which differences were significant. A significance test, if you will. One among them, William Gosset, was more mathematically inclined. And from him came the significance test. All in the interest of better—and more profitable—beer.
在本章中，我们将介绍三位与显著性和假设检验的创立最相关的科学家：戈塞特、罗纳德·费舍尔和杰西·奈曼。他们每个人面对问题时所关注的重点截然不同。戈塞特想要确定酿造最佳啤酒的工艺。他的任务是工程学。费舍尔想要科学知识。他的任务是科学。奈曼想要做出最佳选择。他们各自都提出了使用数据和新近命名的“数理统计”领域在假设之间做出决定的方法。他们对决定的含义以及决定的最终结果的看法截然不同。
IN THIS CHAPTER, we look at three scientists most associated with the creation of significance and hypothesis testing: Gosset, Ronald Fisher, and Jerzy Neyman. Each came to the problems with radically different preoccupations. Gosset wanted to identify the brewing process to yield the best beer. His was an engineering task. Fisher wanted scientific knowledge. His was a scientific task. And Neyman wanted to make the best choices. Each advanced methods for deciding among hypotheses using data and the newly christened field of “mathematical statistics.” They utterly differed on what deciding meant—what deciding added up to.
戈塞特是一位工业统计学家；费舍尔是一位绅士科学家；奈曼是一位 20 世纪中叶的理性规划师。戈塞特发明统计学的目的是通过最大限度地提高品味、一致性和耐用性来最大化利润。费舍尔发明统计学是为了创造关于世界真实情况的科学知识。奈曼则试图根据手头的证据以最理性的方式帮助人们做出选择。他们共同创造了对科学意义的新理解——通过使用统计数据来检验假设。
Gosset was an industrial statistician; Fisher a gentleman scientist; and Neyman a mid-century rational planner. Gosset devised statistics for the purpose of maximizing profits by maximizing taste and consistency and durability. Fisher devised statistics to create scientific knowledge of the way the world really is. Neyman sought to help make choices in the most rational manner according to the evidence at hand. Together they created a new understanding of what it means to be scientific—by testing hypotheses using statistics.
格塞特：啤酒酿造测试
Gossett: Testing for Beer Making
1923 年，一位关键的吉尼斯员工解释了进行科学实验的意义：“测试谷物品种的目的是找出哪种谷物能给农民带来最好的回报。” 2实验很棒，但代价昂贵。今天，我们担心处理过多数据所带来的巨大问题。相比之下，这里的问题在于数据太少，以及进行额外实验以获取更多信息的实际成本。作者威廉·戈塞特 (William Gosset) 是一名工程师和实用数学家，后来成为酿酒师，这个问题的答案决定性地改变了统计学和实验科学。高尔顿和皮尔逊的技术适用于大量数据。酿酒师戈塞特认为，需要一些不同的东西：
In 1923, a key Guinness employee explained the point of doing scientific experimentation: “The object of testing varieties of cereals is to find out which will pay the farmer best.”2 Experiment was great—but expensive. Today, we worry about the enormous problems of dealing with too much data. By contrast, the problem here was too little data and the real costs of doing additional experiments to get more. The author, William Gosset, was an engineer and practical mathematician turned brewer, and the answer to the question transformed the science of statistics—and of experimentation—decisively. The techniques of Galton and Pearson were fine for large sets of data. Something different was needed, the brewer Gosset argued:
有时需要从非常小的样本判断结果的确定性，而这个样本本身就是变异性的唯一指示。一些化学实验、许多生物实验以及大多数农业和大规模实验都属于这一类，迄今为止，这几乎超出了统计调查的范围。3
it is sometimes necessary to judge of the certainty of the results from a very small sample, which itself affords the only indication of the variability. Some chemical, many biological, and most agricultural and large scale experiments belong to this class, which has hitherto been almost outside the range of statistical enquiry.3
卡尔·皮尔逊关注的是大量生物数据，而戈塞特关注的是工业应用，在这些应用中，观测数据量很小。4收集大量关于人的测量数据比进行昂贵的工业和农业实验更容易、更便宜。这促使戈塞特设计出评估基于小数据集的推论可信度的技术，并尽量降低这样做的成本。实验方法需要通过最小化成本来最大化利润。他写信给卡尔·皮尔逊说：“在我们这样的工作中，要达到的确定性程度必须取决于遵循实验结果所获得的经济利益，与新方法增加的成本（如果有的话）和每次实验的成本相比。” 5在没有成本的情况下，无法对重要性做出普遍的判断。在度假期间，戈塞特骑自行车去见皮尔逊，并学习了“当时使用的几乎所有方法” 。6
Whereas Karl Pearson focused on large amounts of biological data, Gosset was concerned with industrial applications where the amount of observed data was small.4 Collecting lots of measurements about people was easier and cheaper than undertaking expensive sets of industrial and agricultural experiments. This led Gosset to devise techniques for assessing the confidence in inferences based on small data sets and to minimize cost in doing so. Methods of experiment needed to maximize profit by minimizing cost. “In such work as ours,” he wrote to Karl Pearson, “the degree of certainty to be aimed at must depend on the pecuniary advantage to be gained by following the result of the experiment, compared with the increased cost of the new method, if any, and the cost of each experiment.”5 No universal judgement about significance could be made in absence of cost. While on vacation, Gosset bicycled to meet Pearson and learn “nearly all the methods then in use.”6
戈塞特测试
GOSSET’S TEST
假设你喜欢喝啤酒。你甚至更喜欢通过酿造啤酒来赚钱。你想提高大麦的产量，因为你酿造了健力士啤酒。你可以对肥料、灌溉、品种等进行不同的实验，但你如何确定你做的事情有效呢？戈塞特设计了一种测试来帮助他评估实验，现在称为学生 t 检验。想象一下，你有十块相邻的田地，你在每隔一块田地里种植两种大麦中的一种。你测量每块田地中后续作物的大小，并将每块大麦类型 1 的田地的大小与每块大麦类型 2 的田地的大小进行比较。几乎总会有一些差异。但这些差异中哪些是由于偶然变化造成的，哪些是由于各种品种造成的？我们需要确定这种差异仅仅是由于正常波动而不是任何特定原因（在本例中是不同品种）造成的可能性有多大。在 Gosset 检验中，我们通过将差异的平均值除以数据点数的平方根的标准差来计算统计数据。然后，我们可以在表格中查找从随机变异中获得具有这些品质的数据的可能性。如果随机变异极不可能产生这些数据所描述的植物产量，那么我们就有充分的理由认为一种大麦生长得更好。后来我们了解到这是在检验一个假设——一种大麦生长得更多——与一个“零”假设——两种大麦生长得差不多——之间的对比。
Suppose you like beer. Even more, you like to make money from brewing beer. You want to improve the yield of barley because you brew Guinness beer. You can perform different experiments with fertilizer, irrigation, varieties, and so forth, but how would you know with certainty that something you did worked? Gosset devised a test to help him assess experiments, now known as Student’s t-test. Imagine you have ten fields next to each other, and you plant one of two kinds of barley in every other one. You measure the size of the subsequent crop in each field, and compare the size of each field of barley type 1 with the size of each field of barley type 2. There will almost always be some difference. But what of the difference is due to chance variations and what’s due to the various varieties? We need to determine how likely the difference is simply due to normal fluctuation and not any particular cause, in this case the different variety. In Gosset’s test, we compute a statistic by dividing the mean of the differences by the standard deviation over the square root of the number of data points. We can then look up on a table how probable it is that we would get data with these qualities from random variations. If it is extremely unlikely that random variations would have produced yields of plants described by this data, then we have good grounds to think that one variety of barley grows better. Later we’d come to know this as testing a hypothesis—that one form of barley grows more—against a “null” hypothesis—that both forms of barley grow about the same.
你可能想知道什么是“极不可能”。1/10 的概率是否足以证明某件事只是随机变化？1/20？1/100？1/1000000？这是一个决定实验者。这是一个选择。它不是由科学给出的。这是一个决定，决定什么能让我们放心，我们有信心一件事是这样的，而另一件事不是。
You’re probably wondering what’s “extremely unlikely.” Is a 1/10 chance good evidence something is just a random variation? 1/20? 1/100? 1/1000000? This is a decision of the experimenter. It’s a choice. It’s not given by science. It’s a decision about what would make us comfortable that we have confidence that one thing is the case, and another not.
对于戈塞特来说，这是为了他的公司赚钱。他有兴趣帮助吉尼斯实现利润最大化，而我们可能获得的确定性水平需要满足这一要求。因此，戈塞特没有提供任何统计显著性规则。
For Gosset, it was about making money for his firm. He was interested in helping Guinness maximize profit, and the level of certainty we might acquire needed to attend to that concern. So Gosset provided no one rule of statistical significance.
通过他的研究，戈塞特提供了一种全新的思考方式，让我们思考如何在不确定的情况下做出决策。他提供了新的数学方法，让我们能够根据不确定的证据选择一种行动方案而不是另一种。他的新数学方法可以帮助我们判断某些证据是否足以证明应用的可行性。戈塞特称这是一种“金钱”方法，用于回答什么才是行动和商业决策所需的充分知识。
With his studies, Gosset provided a radical new way to think about making decisions under conditions of uncertainty. He offered new mathematical techniques for choosing one course of action over another based on inconclusive evidence. His new math helps us decide whether some evidence was conclusive enough for the application in question. Gosset called this a “pecuniary” approach to questions of what constitutes adequate knowledge for action, for making business choices.
通过聘用和表彰戈塞特这样的应用科学家，吉尼斯率先将科学推理应用于啤酒生产以及该过程的所有工业和农业方面。农作物的变异越来越多地让位于品种的标准化。啤酒的本地品质让位于标准化且通常更稳定的啤酒。小样本实验中变异的挑战促使戈塞特探索数学技术。他们的目标不是科学知识本身。戈塞特和他的团队试图使用数学工具优化整个酿造过程，以提高质量、耐用性并最终提高利润。7
By employing and celebrating applied scientists such as Gosset, Guinness pioneered the application of scientific reasoning to the production of beer and all the industrial and agricultural facets of the process. Variation of crops increasingly gave way to standardization of varieties. Local qualities of brews gave way to standardized and usually more stable beers. The challenges of the variation within experiments with small sample sizes motivated Gosset to explore mathematical techniques. The goals were not scientific knowledge itself. Gosset and his team sought to optimize the entire brewing process using mathematical tools, to increase quality, durability, and ultimately profits.7
作为吉尼斯的一名员工，戈塞特必须以假名发表他的研究成果，所有与酿酒有关的引用都用其他主题代替。他的作品现在以“学生”的名义为人所知。正如我们将看到的，统计学和数据科学的故事充满了掩盖事实的作者激发他们探索和创新的数据和动机。
As an employee of Guinness, Gosset was required to publish his results under a pseudonym, with all references to brewing replaced with other subjects. His work is now known under the name “Student.” As we will see, the story of statistics and data science is replete with authors obscuring the data and motives that provoked their inquiries and innovations.
戈塞特的研究成果发表在皮尔逊的顶级统计学和优生学杂志《Biometrika》上，最初并未引起统计学界的广泛关注。统计学界更关注用数字描述社会和自然世界，而不是对有关它们的假设进行分类。随着时间的推移，戈塞特的思想开始重新定位科学和社会科学，这在很大程度上是通过费舍尔和奈曼的研究实现的——他们两人都对戈塞特的研究进行了重大的修改。
Published in Pearson’s premier journal of statistics and eugenics, Biometrika, Gosset’s work initially garnered little interest within the broader statistical community. That community was more focused on describing the social and natural world numerically than in sorting through hypotheses about them. In time, Gosset’s ideas came to reorient the sciences and the social sciences largely through the work of Fisher and then of Neyman—each dramatically reworked Gosset’s work.
与戈塞特一样，费舍尔也致力于将统计学应用于现实世界的问题，尤其是农业生产力问题。与戈塞特不同，费舍尔追求的是科学知识本身。他借鉴了戈塞特在小数据集方面的工作，提出了科学实验本身的革命性概念。
Like Gosset, Fisher worked extensively on applying statistics to real world problems, above all questions of agricultural productivity. Unlike Gosset, Fisher sought scientific knowledge itself. He drew on Gosset’s work on small data sets to offer a revolutionary conception of scientific experiment itself.
统计学家弗洛伦斯·南丁格尔·戴维 (Florence Nightingale David) 在讨论数理统计发展中的伟大人物时解释说，戈塞特“提出问题，[埃贡]皮尔逊 (Egon Pearson) 或费舍尔 (Fisher) 将它们转化为统计语言，然后奈曼开始研究数学。” 8
Discussing the great figures in the development of mathematical statistics, the statistician Florence Nightingale David explained, Gosset “asked the questions and [Egon] Pearson or Fisher put them into statistical language and then Neyman came to work with the mathematics.”8
费舍尔：通过测试来获取真相
Fisher: Testing for Truth Making
1925 年，罗纳德·费舍尔毫不留情地谴责当时统计学的实用性：“传统的统计过程机制完全不适合实际研究的需要。它不仅需要用大炮来打麻雀，而且它还会打不中麻雀！”大炮不适用于小数据集：“建立在无限大样本理论基础上的复杂机制对于简单的实验室数据来说不够准确。”简单的实验室数据需要新技术，像戈塞特的技术一样，以数学水平的提高和对科学实验本身的更好理解。“只有系统地解决小样本问题，才有可能将准确的测试应用于实际数据。” 9
In 1925, Ronald Fisher spared no words in denouncing the usefulness of the statistics of his day: “the traditional machinery of statistical processes is wholly unsuited to the needs of practical research. Not only does it take a cannon to shoot a sparrow, but it misses the sparrow!” The cannon wasn’t good for small sets of data: “The elaborate mechanism built on the theory of infinitely large samples is not accurate enough for simple laboratory data.” Simple laboratory data needed new techniques, techniques like those of Gosset, grounded in better mathematics and a better grasp of scientific experiment itself. “Only by systematically tackling small sample problems on their own merits does it seem possible to apply accurate tests to practical data.”9
费舍尔开发的假设检验形式产生了数十亿个“p 值”，这些值从 20 世纪中叶至今一直主导着大量科学成果，并且是法律要求接受多种形式的医疗和药物治疗的有效性。费舍尔为统计学奠定了新的数学基础，取代了以前对科学的理解。
From the form of hypothesis testing Fisher developed came the billions of “p values” that have dominated much scientific production from the mid-twentieth century to this day and are legally required for accepting the efficacity of many forms of medical and pharmaceutical treatments. Fisher set down a new mathematical basis for statistics as a replacement for previous understandings of what makes up science.
和戈塞特一样，费舍尔也是在实际的农业环境中开发他的工具，就他所研究的领域而言，就是罗瑟姆斯特德实验站。10 1922年，戈塞特在该站遇到了费舍尔。和戈塞特一样，他也为旧的实验项目带来了全新水平的数学复杂性。
Like Gosset, Fisher developed his tools in a practical agricultural context, in his case the Rothamsted Experimental Station.10 Gosset met Fisher at the station in 1922. Like Gosset, he brought a new level of mathematical sophistication to an older experimental program.
在罗瑟姆斯特德，费舍尔获得了几代农业实验积累的数据。他可以自由地帮助指导实验站的实验设计。他的女儿兼传记作者写道：“罗瑟姆斯特德的活动、工作人员的兴趣和问题、喝茶时的讨论，都极大地激发了费舍尔的聪明才智和创造力。” 11一系列与应用农业问题相关的新数学方法的论文迅速出现。他的论文很快首次将重点放在了观察到的变化的“重要性”上。12这些问题通常很难解决。“人们经常认为，”一篇论文开头写道，“不同栽培植物的品种不仅在适应不同气候和土壤条件方面存在差异，而且在对不同肥料的反应方面也存在差异。”根本问题是如何从轶事证据转向更“关于不同肥料相对价值的确凿证据”。13
At Rothamsted, Fisher was presented with data accumulated over generations of agricultural experimentation. And he was at liberty to help guide the design of experiments at the station. “The activities at Rothamsted,” his daughter and biographer wrote, “the interests and the problems of the staff, the discussions over a cup of tea, all were a great stimulus to Fisher’s ingenuity and inventiveness.”11 A stream of papers with new mathematical approaches tied to applied agricultural problems quickly appeared. His papers soon focused on the “significance” of observed variations for the first time.12 The problems could often be, well, shitty to solve. “It is not infrequently assumed,” one paper began, “that varieties of cultivated plants differ not only in their suitability to different climatic and soil conditions, but in their response to different manures.” The fundamental question was how to move from anecdotal evidence to more “conclusive evidence as to the relative value of different manures.”13
答案来自统计测试，这是戈塞特开创的测试类型。费舍尔充分利用自己的数学技能和农业站的农业测试经验，将戈塞特的方法重新改造为科学实验本身的新解释。
The answer came from statistical testing, the sort of testing Gosset had pioneered. Drawing equally upon his skill in math and the experience of agricultural testing at the agricultural station, Fisher recast Gosset’s approach into a new account of scientific experimentation itself.
1925 年，他将自己的方法整合到教科书《研究人员的统计方法》中。这本书广泛传播了他的实验方法。“实验室工作人员每天接触的统计问题激发了纯数学研究，这里介绍的方法就是基于此。” 14这本书果断地从对“总量或平均值”感兴趣的旧统计学转向“研究任何可变现象变化的原因，从小麦产量到人类智力”，这需要“检查和测量出现的变化。” 15
In 1925, he integrated his approaches into a textbook, Statistical Methods for Research Workers. The book spread his approach to experimentation widely. “Daily contact with the statistical problems which present themselves to the laboratory worker has stimulated the purely mathematical researches upon which are based the methods here presented.”14 The book decisively moves from older statistics interested in “aggregate, or average, values” to the “study of the causes of variation of any variable phenomenon, from the yield of wheat to the intellect of man,” which requires “the examination and measurement of the variation which presents itself.”15
创建实验涉及陈述要与零假设进行检验的假设。支持假设的显著结果意味着我们相信我们只能在极小的百分比时间内从零假设中获得实验数据，比如二十分之一，即 5%。Fisher 否认任何通用阈值，但他认为，“实验者通常和方便地将 5% 作为显著性的标准水平，这意味着他们准备忽略所有未达到此标准的结果。” 16
Creating an experiment involved the statement of the hypothesis to be tested against a null hypothesis. A significant result in favor of hypothesis means that we believe that we would only obtain the data from an experiment from a null hypothesis a very small percentage of the time, say one out of twenty, or 5 percent. While denying any universal threshold, Fisher argued, “it is usual and convenient for experimenters to take 5 per cent. As a standard level of significance, in the sense that they are prepared to ignore all results which fail to reach this standard.”16
费舍尔的严谨原则旨在抵御许多偏见、希望和梦想，它们会影响我们思考数据的判断。如何消除所有可能破坏我们努力找出一个潜在原因进行调查的各种潜在原因（无论是可见的还是不可见的）？如何避免科学家通过选择比较来歪曲事实，并排除无数其他原因？费舍尔坚持在实验设计中采用随机化方法，以消除可能使调查陷入混乱的原因。费舍尔坚持对要测试的事物进行随机化，以便“可以保证显著性检验不会因尚未消除的干扰原因而受到破坏”。费舍尔解释说，随机化“使实验者不必担心考虑和估计可能干扰其数据的无数原因的严重程度。” 17为了排除实验后操纵数据所固有的危险，费舍尔要求在开始数据收集之前确定数据分析计划和要测试的假设。例如，在进行药物试验时，我们必须使用（通常是预先登记）某种程序来随机选择哪些患者将接受我们正在测试的药物，哪些患者将接受安慰剂。
Fisher’s austere doctrine was designed to ward off the many biases, the hopes, the dreams that cloud judgment in thinking about data. How to eliminate all the various potential causes, seen and unseen, that might disrupt our efforts to isolate one potential cause for investigation? To avoid the often unconscious ways scientists might tilt the scales by selecting comparisons and to rule out the countless other causes that might muck up an inquiry, Fisher insisted on the need for randomization in the creation of an experiment. Fisher insisted upon randomization of things to be tested so that “the test of significance may be guaranteed against corruption by the causes of disturbance which have not been eliminated.” Randomization, Fisher explained, “relieves the experimenter from the anxiety of considering and estimating the magnitude of the innumerable causes by which his data may be disturbed.”17 To preclude the dangers inherent in manipulating data after an experiment, Fisher required that the plan for the analysis of the data and the hypothesis to be tested be locked into place before data collection began. In doing a trial of a pharmaceutical, for example, we must use, and typically preregister, some procedure to choose randomly which patients will receive the drug we’re testing, and which patients will receive a placebo.
尽管费舍尔的许多要求至今仍被认为是良好科学实践的必要条件，但他的其他要求仍然引起人们的恐慌。从他那个时代开始，批评者就一直认为随机化充其量是浪费，最坏的情况是不道德和致命的。在制药业中，随机对照试验 (RCT) 的黄金标准无疑保护了消费者免受负面副作用和无效药物的侵害，但代价是放慢了药物审批速度，缩小了认定疗法有效的依据。长期以来，业界一直谴责实验性治疗进入市场的滞后，这在 20 世纪 80 年代和 90 年代成为识别和治疗 HIV 感染运动的集结点。18早在那之前，戈塞特就曾试图让费舍尔承认随机试验所需的低效率，但没有成功。19
While many of Fisher’s demands echo as essential to good scientific practice to this day, other of his demands still cause consternation. Critics from his time onward have challenged randomization as wasteful at best and unethical and deadly at worst. In pharmaceuticals, the gold standard of randomized controlled trials (RCTs) has unquestionably protected consumers from negative side effects and ineffective drugs, but at the cost of slowing drug approval and of narrowing the grounds for deeming therapies effective. The lag in experimental treatments coming to market, long deplored by industry, became a rallying point in the movement to recognize and treat HIV infections in the 1980s and 1990s.18 Long before that, Gosset tried in vain to get Fisher to acknowledge the inefficiencies required for randomized trials.19
自由、优生学和种族提升
LIBERTY, EUGENICS, AND THE UPLIFT OF RACES
戈塞特想要一种更好的啤酒原料测试程序。费舍尔则希望提高通过基于实验的知识实现人类自由。“然而，只要人类智力的自由只能是按照既定的教条数据得出结果，而无法获得只有直接观察才能获得的未知真理，那么人类智力的解放就必定是不完整的。” 20只有经验知识才能克服教条。长期以来，为实验程序辩护一直困扰着哲学家：我们如何从个人经验中得出概括？费舍尔在法西斯主义在欧洲蔓延时写道，人类自由需要实验：“实验设计的艺术和对实验结果的有效解释，只要它们在技术上可以完善，就必须构成行使充分知识自由这一主张的核心。” 21
Gosset wanted a better procedure for testing ingredients to make beer. Fisher sought nothing less than to enhance human freedom through knowledge based on experiment. “The liberation of the human intellect must, however, remain incomplete so long as it is free only to work out the consequences of a prescribed body of dogmatic data, and is denied the access to unsuspected truths, which only direct observation can give.”20 Only experiential knowledge could overcome dogma. Justifying experimental procedure has long bedeviled philosophers: how can we come to generalizations from individual experiences? Writing as fascism was spreading across Europe, Fisher argued that human freedom required experiment: “the arts of experimental design and of the valid interpretation of experimental results, in so far as they can be technically perfected, must constitute the core of this claim to the exercise of full intellectual liberty.”21
对于费舍尔来说，科学并不是一项提高利润的机械事业。对他来说，人类进步“不仅仅是生产出一台高效的工业机器，或一个反面美德的典范，而是激发人类所有独特的特征，所有人类最优秀的品质，所有不同的品质，有些很明显，有些则非常微妙，但我们认为这些都是人类的优秀品质。” 22他的愿景不是全人类的提升，而是人类不同种族之间的冲突。在一篇年轻时的文章中，费舍尔解释说：“未来分布广泛、硕果累累、成功的种族属于当今的主导国家；而国家之所以能占据主导地位，主要取决于组成它们的人的忠诚、进取和合作能力。” 23费舍尔看到了民族间冲突的加剧，但他并没有从工业的角度来看待人类进步。费舍尔的优生学是一锅令人陶醉的大杂烩，是达尔文、尼采和英国国教的不可思议的结合。一场种族战争正在进行中，但他坚持认为，这不是一场工业种族战争。
For Fisher, science was not a mechanical enterprise of improving profit. For him, human progress involved “not a question merely of producing a highly efficient industrial machine, or a paragon of the negative virtues, but of quickening all the distinctively human features, all that is best in men, all the different qualities, some obvious, some infinitely subtle, which we recognise as humanly excellent.”22 His vision was less one of universal uplift than conflict among the different races of humanity. In one youthful piece, Fisher explained, “The widespread, fruitful, and successful races of the future belong to the dominant nations of to-day; and nations are rendered dominant principally by the loyalty, enterprise and cooperative ability of the people who compose them.”23 Fisher saw a heightening of conflict among national races but did not cast human progress in industrial terms. Fisher’s eugenics was a heady stew, an unlikely combination of Darwin, Nietzsche, and Anglicanism. A race war was afoot, but it was, he maintained, not an industrial race war.
伟大文明衰落的幽灵萦绕在费舍尔的生物学工作。和其他优生学家一样，他感觉到人类生殖出现了惊人的逆转。在经济发达、市场驱动的文明中，经济和文化上最成功的人生殖成功率最低；他认为，社会上最优秀的人注定无法充分繁殖，将被较差的人所淹没。从罗马到当代英国，文明社会的经济关系都是不利于人类的。而最优秀的人的逐渐消失，无情地导致文明失去了人类文化的高级装备，并最终导致这些文明和创造它的种族的衰落。
Ghosts of the decline of great civilizations haunt Fisher’s biological work. Like other eugenicists, he sensed a striking inversion of human reproduction. In economically advanced, market-driven civilizations, the most financially and culturally successful were the least reproductively successful; the best of society, he argued, were slated not to reproduce adequately, and would be swamped by lesser human beings. The economic relations of civilized societies, from Rome to contemporary Britain, were dysgenic. And the progressive elimination of the best people led inexorably to civilizations losing the higher accoutrements of human culture and eventually to the decline of those civilizations and the races that had created it.
换句话说，让经济逻辑占主导地位就意味着失去人类最优秀的品质，包括最高级的文化形式，包括科学，以及基因最优越的人。考虑到他的优生学框架，费舍尔对那些设想按照经济效率进行假设检验的统计学对手做出愤怒的反应并不奇怪。他们把反对教条主义的堡垒变成了一种低级的文化形式。
In other words, to permit an economic logic to dominate was to lose the best of humanity, both the highest forms of culture, including science, and the most genetically superior people. Given his eugenical framework, it is hardly surprising that Fisher reacted angrily to statistical rivals who envisioned hypotheses testing along the lines of economic efficiency. They were turning the bulwark against dogmatism into a lesser cultural form.
针对成本函数
AGAINST COST FUNCTIONS
正如我们所见，戈塞特根据潜在利润来评估实验的有效性。这种务实理解的后续版本让费舍尔感到震惊，他开始认为这是以低劣的工业动机侵犯科学探究的纯粹性。费舍尔解释说，没有任何金钱价值可以决定知识的轮廓：
As we saw, Gosset evaluated the efficacy of an experiment in terms dictated by potential profit. Later versions of this pragmatic understanding horrified Fisher, who came to understand it as violating the purity of scientific inquiry with base industrial motives. Fisher explained that no pecuniary value could decide the contours of knowledge:
在归纳推理中，我们不引入错误判断的成本函数，因为在科学研究中，人们认识到，今年而不是以后取得或未能取得某项科学进步，都会对以下方面产生影响：研究计划，以及科学知识的有利应用，这些是无法预见的。……我们不试图评估这些后果，也不认为它们能够以任何形式的货币进行评估。24
in inductive inference we introduce no cost functions for faulty judgements, for it is recognized in scientific research that the attainment of, or failure to attain to, a particular scientific advance this year rather than later, has consequences, both to the research programme, and to advantageous applications of scientific knowledge, which cannot be foreseen. . . . We make no attempt to evaluate these consequences, and do not assume that they are capable of evaluation in any sort of currency.24
费舍尔最大的竞争对手却不同意这一观点。统计学不应执着于追求科学真理，而应专注于做出选择——无论是在商业领域还是在科学领域。
Fisher’s greatest rival disagreed. Rather than pining after scientific truth, statistics needed to focus on making choices— in business as well as in science.
奈曼：决策测试
Neyman: Testing for Decision-making
波兰数学家耶日·奈曼几十年来一直认为，大多数假设检验的问题在于，大多数人认为它与真理有关。奈曼认为它与选择有关。“我们不希望知道每个假设是真还是假，而是寻找规则来规范我们对它们的行为，遵循这些规则，我们就能确保在长期经验中，我们不会经常犯错。” 25我们需要更有效的测试，而不是真理：“任何基于概率论的测试本身都不能提供任何有价值的证据来证明该假设是真还是假。”
The problem with most hypothesis testing, the Polish mathematician Jerzy Neyman argued for decades, is that most people thought it was about truth. Neyman argued it was about choices. “Without hoping to know whether each separate hypothesis is true or false, we may search for rules to govern our behaviour with regard to them, in following which we insure that, in the long run of experience, we shall not be too often wrong.”25 We needed more efficient tests, not the truth: “no test based upon the theory of probability can by itself provide any valuable evidence of the truth or falsehood of that hypothesis.”
奈曼和他的同事卡尔之子埃贡·皮尔逊认为，费舍尔未能意识到假设检验中的第二个危险。费舍尔担心接受一个错误的假设；奈曼和皮尔逊强调，我们需要担心拒绝一个我们应该接受为真的假设。因此，在检验假设时，我们需要平衡两种类型的错误，即 I 型和 II 型错误。费舍尔建议始终用零假设来检验假设。奈曼和皮尔逊坚持需要比较相互竞争的假设。
Neyman and his collaborator Egon Pearson, son of Karl, argued that Fisher had failed to appreciate a second danger in hypothesis testing. Fisher worried about accepting a hypothesis that was false; Neyman and Pearson stressed the need to worry about rejecting a hypothesis that we should accept as true. In testing hypotheses, then, we need to balance two types of error, quickly branded type I and type II. Fisher advised always testing a hypothesis against a null hypothesis. Neyman and Pearson insisted on the need to compare competing hypotheses.
这种截然不同的统计方法如何测试出现了吗？奈曼将极其深奥的数学、对知识的怀疑态度和实际的农业工作结合在一起。统计并不是奈曼的明显职业。奈曼是一名生活在战时俄罗斯的波兰学生，在学习实验物理时遭遇惨败后，他开始致力于在高度抽象的理论基础上重塑数学。这种数学似乎与所有实际应用都相距甚远，甚至与理论物理也相距甚远。当时和现在一样，对于攀登抽象冰坡的纯数学家来说，工作机会很少，不久之后，奈曼就发现自己从事高度应用的统计工作，以支付账单和提供住房。
How did this dramatically different approach to statistical testing arise? Neyman brought together extremely recondite mathematics, a skeptical vision of knowledge, and practical agricultural work. Statistics was not Neyman’s obvious vocation. After failing spectacularly as a student of experimental physics, Neyman, a Polish student living in wartime Russia, became caught up in the effort to recast mathematics on highly abstract theoretical grounds. This mathematics seemed distant from all practical application, even theoretical physics. Jobs were few, then as now, for pure mathematicians climbing the icy slopes of abstraction, and before long Neyman found himself working in highly applied statistics jobs to pay the bills and to provide housing.
他的观点是在 20 世纪 20 年代新独立（且暂时独立）的波兰的实验性农业工作中形成的。奈曼通过实验和当时最先进的理论数学，参与了科学史学家西奥多拉·德赖尔所说的“一场将主权波兰想象成一个现代繁荣的农业民族国家的动态运动”。26与费舍尔一样，奈曼和他的同事在学生的论文中找到了分析农业实验的强大工具。奈曼的方法应该在理性创造繁荣经济的梦想中理解。
His views emerged amid the experimental agricultural work in newly—and temporarily—independent Poland in the 1920s. Neyman was working within what the science historian Theodora Dryer calls “a dynamic movement to imagine sovereign Poland as a modern and prosperous agrarian nation state” through experimentation and the most sophisticated theoretical mathematics of the day.26 Like Fisher, Neyman and his colleagues found powerful tools for analyzing agricultural experiments in Student’s papers. Neyman’s approach should be understood within a dream of rational creation of a thriving economy.
他是如何将自己钟爱的高度抽象的数学与具体的农业工作联系起来的？奈曼借鉴了他最喜欢的一本书《科学语法》，作者是英国优生学家和统计学家卡尔·皮尔逊。27他在书中吸收了皮尔逊的科学观点，即科学“以毫不妥协的方式攻击各种权威”，抛弃所有现存的教条，无论是宗教的、社会的还是科学的——对于正准备推翻沙皇和教会的俄罗斯年轻人来说，这些都是令人兴奋的东西。他接受了皮尔逊对我们真正知道的东西的深刻怀疑。晚年，奈曼解释说，
How did he connect his beloved highly abstract mathematics with this concrete agricultural work? Neyman drew upon a favorite book, Grammar of Science, by the English eugenicist and statistician Karl Pearson.27 There he imbibed Pearson’s vision of science attacking, Neyman said, “in an uncompromising manner all sorts of authorities,” throwing off all existing dogma, whether religious, social, or scientific—heady stuff for young people in Russia on the verge of overthrowing tsars and church alike. He embraced Pearson’s profound skepticism about what we really know. Later in life, Neyman explained,
我最喜欢的一个观点是从马赫那里通过卡尔·皮尔逊的《科学语法》学到的，那就是科学理论不过是自然现象的模型，而且往往是不充分的模型。模型是一组关于虚构实体的虚构假设，如果将这些虚构实体视为所研究现象的适当元素的表示，那么构成模型的假设的结果预计会与观察结果一致。如果在所有相关试验中，一致性程度都令人满意，那么我们认为该模型是一个充分的模型。28
One of my favorite ideas, learned from Mach via Karl Pearson’s “Grammar of Science”, is that scientific theories are no more than models of natural phenomena, frequently inadequate models. A model is a set of invented assumptions regarding invented entities such that, if one treats these invented entities as representations of appropriate elements of the phenomena studied, the consequences of the hypotheses constituting the model are expected to agree with observations. If, in all relevant trials, the degree of conformity appears to us satisfactory, then we consider the model an adequate model.28
对于皮尔逊来说，知识总是暂时的。他解释说：“信念应被视为知识的附属物：在需要决策但可能性又不至于大到足以构成知识的情况下，信念是行动的指南。” 29我们所能做的最好的事情就是坚信一个模型最符合我们手头的现象。我们对周围事物的真正原因和运作方式没有任何真正的洞察力。奈曼的任务是展示他所崇拜的高度抽象的数学如何帮助评估和构建模型。
For Pearson, knowledge was always provisional. “Belief,” he explains, is “to be looked upon as an adjunct to knowledge: as a guide to action where decision is needful, but the probability is not so overwhelming as to amount to knowledge.”29 The best we can do is assert confidence in one model as best conforming to the phenomena we have at hand. We don’t have any real insight into the true causes and things at work around us. Neyman’s task was to show how the highly abstract mathematics he adored can help evaluate and construct models.
这位年轻学者一直资金短缺，因此奈曼担任过一系列应用统计职位，直到他获得资金前往英国与他的老偶像卡尔·皮尔逊一起工作。令他惊讶的是，皮尔逊对新抽象数学知之甚少，但他帮助奈曼获得了巴黎奖学金，奈曼在那里主要回到了抽象数学的世界。
Money was always short for the young scholar, so Neyman held a series of applied statistical positions, until he received funding to travel to England to work with Karl Pearson, his old hero. To his surprise, Pearson knew little of the new abstract mathematics, but helped Neyman secure a fellowship to Paris, where Neyman returned largely to the world of abstract mathematics.
戈塞特又回到我们的故事中。他再次启发了那些更倾向于数学的统计学家。1926 年，戈塞特写信给皮尔逊的儿子埃贡，提出了一系列关于假设检验意义的问题。
Here Gosset returns to our story. Again he inspired the more mathematically inclined statisticians. In 1926 Gosset wrote to Pearson’s son Egon with a series of questions about the meaning of hypothesis testing.
如果存在任何其他假设，能够以更合理的概率解释样本的出现，比如 0.05……，那么你将更倾向于认为原始假设不正确。30
If there is any alternative hypothesis which will explain the occurrence of the sample with a more reasonable probability, say .05 . . . , you will be very much more inclined to consider that the original hypothesis is not true.30
这封信启发了埃贡·皮尔逊在巴黎给奈曼写信，让他重新回到统计学的圈子。这段话包含了奈曼和皮尔逊为费舍尔的测试和科学知识概念所创造的激进替代方案的“思想萌芽” 。31
And this letter inspired Egon Pearson to write Neyman in Paris, bringing him back into the statistical fold. This paragraph contains “the germ of that idea” of the radical alternative that Neyman and Pearson created to Fisher’s conception of testing and of scientific knowledge.31
奈曼主张寻找理由来代替真理，以追求一组行动，而不是另一组行动。奈曼写道：“决定肯定”科学的事物，“并不意味着知道甚至相信。”相反，“这是一种意志行为，先于一些经验和演绎推理，就像人们购买人寿保险一样，即使我们预计自己能活很长时间，我们也会这样做。” 32统计学家埃里希·莱曼 (Erich L. Lehmann) 指出了这一新观点的重大意义。“它首次指出统计理论的目标是系统地寻找最佳程序。在接下来的几十年里，许多理论的发展都是朝着这个目标进行的。” 33
In place of truth, Neyman advocated looking for reasons to pursue one set of actions, and not another. “Deciding to affirm” something scientific, Neyman wrote, “doesn’t mean knowing or even believing.” Rather, “it’s an act of will preceded by some experience and deductive reasoning, just as one takes out life insurance, which we do even if we expect to live for a long time.”32 The statistician Erich L. Lehmann noted the dramatic significance of this new point of view. “For the first time it states the aim of statistical theory to be the systematic search for optimal procedures. Much of the theory developed during the next decades was directed toward this end.”33
奈曼的观点令费舍尔深恶痛绝。34费舍尔曾试图使用统计学的新工具来解释归纳知识是如何实现的。奈曼用它们来否认这种知识的存在，而主张基于证据做出决策：他说，我们最多获得的不是归纳知识，而是基于证据的“归纳行为”。35费舍尔和奈曼（以及他的同事皮尔逊）将在接下来的三十年里展开斗争。他们的论点往往晦涩难懂，围绕着数学是否足以解决人类知识和人类行为的问题。
Neyman’s views were anathema to Fisher.34 Fisher had tried to use the new tools of statistics to explain how inductive knowledge was possible. Neyman used them to deny the existence of such knowledge, in favor of making decisions based on evidence: at best, we don’t gain inductive knowledge, he said, but rather an “inductive comportment” on the basis of evidence.35 Fisher and Neyman (and his colleague Pearson) would fight for the next thirty years. Often arcane in appearance, their arguments revolved around the adequacy of mathematics for resolving questions of human knowledge and human behavior.
费舍尔认为奈曼误解了过去两百年间科学大发展的根源：“自 17 世纪伟大的法国数学家以来，西欧数学思想不断发展，在我们这个时代，通过与自然科学的相互影响，我们终于获得了成果，为归纳推理提供了正确使用的模型，就像欧几里得为演绎逻辑提供了模型一样。” 36但错误更深。对科学的误解意味着奈曼及其众多追随者误解了知识如何让你自由，从而成为了极权主义的盟友。
Fisher argued that Neyman misunderstood what had allowed the great developments of science during the previous two hundred years: “the continuous development of mathematical thought in Western Europe from the great French mathematicians of the 17th century onward, has come to fruition in our own time, by cross-fertilization with the Natural Sciences, in supplying just such a model of the correct use of inductive reasoning, as was supplied by Euclid for deductive logic.”36 But the errors were deeper. Misunderstanding science meant that Neyman and his legions of followers misunderstand how knowledge can set you free, and thereby had become allies of nothing less than totalitarianism.
对于在早期自由思想氛围中长大的人来说，以“推理，严格地说，不能应用于经验数据以得出现实世界中有效的推论”这一学说为代表的意识形态运动，是相当令人恐惧的。不可否认的是，我们西方人认为理所当然的思想自由，现在在地球表面的大部分地区都被成功地否定了。因此，在今天，我们仍然敢于得出自己的结论的逻辑步骤的有效性，不能被过分清楚地阐述，也不能被过分强烈地肯定。37
To one brought up in the free intellectual atmosphere of an earlier time there is something rather horrifying in the ideological movement represented by the doctrine that reasoning, properly speaking, cannot be applied to empirical data to lead to inferences valid in the real world. It is undeniable that the intellectual freedom that we in the West have taken for granted is now successfully denied over a great part of the earth’s surface. The validity of the logical steps by which we can still dare to draw our own conclusions cannot therefore, in these days, be too clearly expounded, or too strongly affirmed.37
真相算法：食谱和“p 值学”
The Truth Algorithm: Cookbookery and “p-value-ology”
第二次世界大战后，所有这些统计工作都产生了两个截然不同的遗产，它们朝着不同的方向发展。第一个遗产在 20 世纪下半叶彻底颠覆了科学的意义，事实上，科学的意义也发生了变化。第二个遗产则导致了统计学家的职业化。追求奈曼风格深奥数学的严谨性，通常与日常使用中存在问题的现实世界数据相距甚远。
In the aftermath of World War II, two distinct legacies of all this statistical effort pulled in divergent directions. The first led to the dramatic upending of what it meant to do science, indeed to be a science, in the second half of the twentieth century. The second led toward professionalized statisticians pursuing the rigor of abstruse mathematics in the style of Neyman, often distant from everyday uses of problematic real-world data.
为了从统计学角度理解实验结果，以及利用数据来判断相互竞争的假设，人们进行了一系列的斗争，这些斗争产生了深远的影响，至今仍影响着我们的世界和我们的感知。最明显的影响是，在我们对“偶然”事件的思考中，统计显著性无处不在，如果我们能确定一个概率（俗称“p 值”）低于 0.05 这个神奇的数字，那么从算法上理解，结果就是正确的。
The fights to understand experimental results statistically, and to use data to adjudicate between competing hypotheses, had lasting impact that still shapes our world and our sensemaking today. The most visible impact is the ubiquity of statistical significance in our thinking about “chance” events, with the algorithmic understanding of a result being true if we can establish a probability—known colloquially as the “p value”—below the magic number of .05.
需要明确的是，这种用算法来设定真相的方法对费舍尔、奈曼和皮尔逊来说都是令人厌恶的。38然而，在对客观确定性和理性决策日益增长的需求面前，他们论证的权力无法经受住时间的考验。在二十世纪下半叶，寻找在零模型下不可能出现的影响成为发表论文、批准药物的标准，在更受欢迎的讨论中，也成为区分偶然性和因果关系的标准。当费舍尔首次发表他关于实验设计的引人注目的新论述时，批评声四起。随着时间的推移，通过更易于阅读的教科书，假设检验成为各种科学的核心。39科学史学家克里斯托弗·菲利普斯解释说：“食品科学家、心理学家、社会学家和医生……认为统计方法提供了一种现成的技术，可以在不可避免的主观环境中做出可靠的因果判断。” 40尽管费舍尔本人强烈反对按部就班地进行实验，但 0.05 是确定显著性的良好阈值这一建议在 20 世纪下半叶成为了可发表结果与科学垃圾之间的分界线。它提供了一种虚假的客观感。
To be clear, such an algorithmic approach to setting truth would have been anathema to Fisher and to Neyman and Pearson alike.38 The force of their arguments, however, was unable to outlive them in the face of the growing demand for objective certainty and for rational decision making. Over the course of the second half of the twentieth century, searching for effects improbable under a null model became the criterion for publication, for approval of drugs, and in more popular discussion, for separating chance from causation. When Fisher first published his dramatic new account of experimental design, critics abounded. In time, through the agency of easier-to-read textbooks, hypothesis testing became central to a wide array of sciences.39 Science historian Christopher Phillips has explained, “food scientists, psychologists, sociologists, and physicians . . . saw statistical methods as providing an off-the-shelf technique to make reliable causal judgments in inescapably subjective settings.”40 While Fisher himself inveighed against a cookbook approach to experiment, the suggestion that .05 was a good threshold for determining significance became essentially the line between a publishable result and scientific garbage in the second half of the twentieth century. It provided a false sense of objectivity.
在一个又一个的领域，这些新的量化方法颠覆了专家的定义以及某一领域的专业知识意味着什么，这些观念发生了改变。在药物功效研究方面，这种转变最为显著和重大。随机试验颠覆了医生判断药物功效的权威。
In field after field, these new quantitative approaches disrupted older visions of who was an expert and what expertise in a field meant. Nowhere was the shift so dramatic and significant as in the investigation of the efficacy of pharmaceuticals. The randomized trial upended the authority of physicians in judging that efficacy.
1961 年，美国医学协会 (American Medical Association) 否认除执业医师以外的任何人有权对治疗的有效性发表意见：“唯一可能最终决定药物疗效和最终用途的方法是大量医学专业人士长期广泛使用该药物。” 41医生和药剂师在 20 世纪中叶一直抵制失去对药物疗效问题的控制权，直到 1962 年美国食品药品管理局 (FDA) 获得重大新权力后，这一局面才急剧恶化。西奥多·波特 (Theodore Porter) 解释说，监管者“认为医生的专业知识不足以控制药品制造商的大胆声明。替代方案是更集中的决策过程，主要基于书面信息。” 42随着 1962 年的《基福弗-哈里斯修正案》 (Kefauver-Harris Amendment) 的出台，随机对照试验成为衡量药物疗效的基准，成为授权药物和记录其副作用的黄金标准。该立法使 FDA 能够对未来的药品进行评估，并回顾已经上市的药品，并下架 1938 年至 1962 年期间批准的危险或无用的药品。
In 1961, the American Medical Association denied that anyone other than practicing physicians should opine on the utility of a treatment: “the only possible final determination as to the efficacy and ultimate use of a drug is the extensive clinical use of that drug by large numbers of the medical profession over a long period of time.”41 Physicians and pharmacists resisted the loss of their control over questions about the efficacy of drugs through the middle of the twentieth century, before dramatically losing ground in 1962, when the Food and Drug Administration (FDA) acquired dramatic new powers. Regulators, Theodore Porter explains, “considered that the expertise of doctors provided an inadequate control on the bold claims of drug manufacturers. The alternative was a more centralized decision process, to be based mainly on written information.”42 With the 1962 law, called the Kefauver-Harris Amendment, the randomized controlled trial became the benchmark for gauging the efficacy of medications, to become the gold standard for authorization of drugs and the documentation of their side effects. The legislation enabled the FDA to gauge drugs going forward and to look retrospectively at medicines already on the market and to remove from sale dangerous or useless drugs previously approved between 1938 and 1962.
那么费舍尔和奈曼/皮尔逊之间的争论又如何呢？很少有人关心应用假设检验的哲学细节。当事物得到广泛应用时，深奥的哲学争论往往会消失。假设检验也是如此。“费舍尔的显著性检验理论……与奈曼-皮尔逊理论中的概念相结合并将其作为“统计数据”本身来教授……毋庸置疑，无论是费舍尔还是奈曼和皮尔逊都不会看好这个他们被迫结婚所生的孩子。” 43
And what of the debates of Fisher and Neyman/Pearson? Few cared about philosophical niceties in applying hypothesis testing. Abstruse philosophical debates tend to disappear when things are given wide application. And so it was with hypothesis testing. “Fisher’s theory of significance testing . . . was merged with concepts from the Neyman-Pearson theory and taught as ‘statistics’ per se . . . it goes without saying that neither Fisher nor Neyman and Pearson would have looked with favor on this offspring of their forced marriage.”43
数学，而非数据：二战后统计学的定位
Mathematics, Not Data: The Placement of Statistics after World War II
随着美国加入第二次世界大战，统计学在战争工作中的应用也出现了井喷式增长，主要集中在纽约的哥伦比亚大学、新泽西的普林斯顿大学和加利福尼亚的伯克利大学。美国人口普查局的 W. Edwards Deming 写道：“统计学家唯一有用的功能就是做出预测，从而为行动提供依据。” 44从如何最好地设置近炸引信，到工厂的质量控制，再到鱼雷的最佳发射角度，这些统计团体在战时取得了无数成功。高度应用的统计学促进了实验分析新方法的产生，其中最重要的是一种称为序贯分析的测试形式。序贯分析将贝尔电话实验室的质量控制程序与戈塞特所推崇的经济主义测试方法统一起来。45
With the entry of the United States into World War II came an explosion of statistics applied to war work, centered at Columbia in New York, Princeton in New Jersey, and Berkeley in California. “The only useful function of a statistician,” wrote W. Edwards Deming of the US Census Bureau, “is to make predictions, and thus to provide a basis for action.”44 Ranging from how best to set proximity fuses, to quality control in factories, to the best angles for torpedoes, the wartime successes of these statistics groups were legion. Highly applied statistics spurred the creation of new approaches to the analysis of experiments, above all a form of testing called sequential analysis. Sequential analysis unified quality control procedures from Bell Telephone Laboratories with the economistic approach to testing like that Gosset celebrated.45
战前，一些统计学家曾试图用纯数学的习语和程序重塑他们的方法，并鼓动将数理统计与更受数据驱动的应用学科区分开来。具有讽刺意味的是，战时应用统计学的成功证明了这种抽象化的趋势是合理的。海军研究办公室 1946 年的一份文件解释说，“战争的需要推动了基础研究，导致哥伦比亚大学形成了新的序贯分析理论。” 46哥伦比亚统计学家哈罗德·霍特林等关键人物利用所有这些成功来证明对高度理论化和高度数学化的统计学的支持是合理的；一位杰出的数学家和计划行政官米娜·里斯在政府内部支持他们。他们在战时取得的成功表明，理论成就了他们的成功。战后不久，在科学似乎不再能获得政府主要资助的时候，海军研究办公室接受了这个故事。“数学进步是科学进步的基础，这一点得到了普遍认可；但在第二次世界大战期间，这一点得到了有力的证明。” 47这样一来，它就腾出了充足的军事资金用于极其理论化的统计学。
Before the war, a few statisticians had attempted to recast their approaches in the idioms and procedures of pure mathematics and agitated for the cordoning off of mathematical statistics from its more data-driven applied cousins. Ironically, wartime successes with applied statistics came to justify this move toward abstraction. The “needs of the war,” a 1946 document from the Office of Naval Research explained, “gave impetus to basic research which resulted in the formation, at Columbia University, of the new theory of Sequential Analysis.”46 Key figures such as the Columbia statistician Harold Hotelling leveraged all these successes to justify the support for highly theoretical and highly mathematical statistics; a remarkable mathematician and program administrator Mina Rees supported them from within government. The stories of their wartime successes explained that theory made their successes possible. Immediately after the war, at a moment when it seemed that science might no longer receive major government funds, the Office of Naval Research accepted this story. “That progress in mathematics is basic to progress in science is generally recognized; but it was forcefully demonstrated during World War II.”47 In doing so, it freed ample military funds for extremely theoretical statistics.
结果，战争期间高度以数据为中心的工作重点发生了变化。奈曼在 20 世纪 40 年代末写道：“二战期间，大多数统计学家都在研究国防问题，这些问题往往具有直接的实际意义。”他召开了一次大型研讨会“以鼓励人们回归理论研究。” 48霍特林解释了应用如何吸引统计学家，但也腐蚀了他们：“应用的召唤是诱人的，它让许多年轻学者放弃了统计理论的培养。” 49抽象在数学中风靡一时，它同样吸引了以奈曼模式工作的统计学家。
As a result, the highly data-focused work of the war shifted in emphasis. “During World War II,” Neyman wrote in the late 1940s, “the majority of statisticians were working on problems of defense which frequently bore the imprint of immediate practical importance.” He held a major symposium “to stimulate the return to theoretical research.”48 Hotelling explained how applications beckoned—but also corrupted—statisticians: “the call of application is enticing, and has led many young scholars to forsake the cultivation of statistical theory.”49 Abstraction was all the rage in mathematics, and it appealed likewise to statisticians working in Neyman’s mode.
奈曼在伯克利的职业生涯不仅展现了数理统计学的巨大进步，还展示了数理统计学作为一门学科的定位。需要明确的是，这并非完全是学术上的探索：奈曼还希望伯克利认识到他的团队应该是一个完整的部门，而不仅仅是数学系内的一个“实验室”。这样做需要奈曼确立数理统计学的数学资质——表明该领域在智力和数学上都足够严谨，值得设立一个部门。
Neyman’s career at Berkeley exhibited tremendous advances in mathematical statistics, but also in defining mathematical statistics as a discipline. To be clear, this was not entirely an academic quest: Neyman also wanted Berkeley to recognize that his group should be a full department, not merely a “lab” within the department of mathematics. Doing so required Neyman to establish the mathematical bona fides of mathematical statistics—to show that the field was sufficiently rigorous, intellectually and mathematically, to warrant a department.
回想起来，数据和我们日常生活之间的关系是否取决于底层的数学性尚不明确分析。然而，当时，数理统计作为一门职业越来越符合纯数学的严谨性和公理化方法。与海军研究办公室一样，新成立的国家科学基金会 (NSF) 也接受了理论数理统计使战时应用成功成为可能的观点，并据此资助了统计。自 1951 年 NSF 成立以来，统计学一直被定位为数学：不是工程学的一个方面，正如人们从战后活动或 21 世纪的经济影响中得出的结论；也不是 Fisher 可能更喜欢的“自然科学”。因此，作为知识领域的学术命脉，资金在很大程度上依赖于建立足够的数学性，推动该领域走向数学概率学家 Leo Breiman 在 2004 年批评的“过度数学化”。1962 年，普林斯顿大学拓扑学家出身的统计学家 John Tukey 认为“数据分析师”应该“使用数学论证和数学结果作为判断的基础，而不是证明的基础或有效性标志”——这有力地表明了数学已经渗透到统计学——或者扭曲了统计学的程度。50 正如我们将在以下章节中看到的那样，这是一个有趣的反事实历史，值得想象：如果美国的学术统计学从二战到本世纪末没有变得如此数学化，模式发现会诞生于电气工程领域，机器学习会诞生于计算机科学领域，数据科学会诞生于工业领域吗？
In retrospect, it’s not clear that the relationship between data and our daily lives hinges on how mathy is the underlying analysis. At the time, however, mathematical statistics as a profession aligned itself ever more with the rigor and the axiomatic approach associated with pure mathematics. Like the Office of Naval Research, the new National Science Foundation (NSF) accepted the view that theoretical, mathematical statistics had made the wartime applied successes possible, and funded statistics accordingly. Since the founding of NSF in 1951, statistics has been located as mathematics: not as an aspect of engineering, as one might conclude from postwar activities or the economic impact in the twenty-first century; nor in the “natural sciences” as Fisher would likely have preferred. The funding, therefore—the academic lifeblood of an intellectual field—has relied largely on establishing sufficient mathiness, driving the field toward what the mathematical probabilist Leo Breiman critiqued as “over-mathematization” in 2004. By 1962, the Princeton topologist-turned-statistician John Tukey argued that “data analysts” ought to “use mathematical argument and mathematical results as bases for judgment rather than as bases for proof or stamps of validity”—a strong sign of how far mathematics had permeated—or warped—statistics.50 As we will see in the following chapters, it’s an entertaining counterfactual history to imagine: Would pattern discovery have been born in electrical engineering, or machine learning in computer science, or data science in industry, had American academic statistics not become quite so mathematical from WWII until the end of the century?
20 世纪 60 年代初，密码学家胡安妮塔·穆迪 (Juanita Moody) 曾对她的雇主——极度神秘的国家安全局 (National Security Agency) 感到遗憾，因为该局无法将其庞大的数据分析能力应用于非机密领域：“我一直担心，我们拥有强大的计算机化能力，但发展速度超乎你的想象，而整个医学界都需要它。” 她一离开国家安全局就说：“我打算自愿去做一些事情，帮助医学界进行计算机化数据处理。你知道这是一个问题，但我们所做的一切都是机密的。” 更糟糕的是，事情本不必这样：“我知道这不必保密，但事实却是如此。”1在 20 世纪 90 年代和 21 世纪大数据兴起的几十年前，国家安全局就已经将数据收集、算法和分析形式制度化。
In the early 1960s, the cryptographer Juanita Moody regretted that her employer, the hypersecretive National Security Agency, could not put its massive capacity for analyzing data to good use in the nonclassified world: “it always worried me that we had great computerized capability just moving faster than you could imagine and that there was this whole, big medical world out there that needed it.” As soon as she could leave the NSA, she said, “I’m going to go volunteer to do something to help the medical world with computerized data processing. You just knew that was a problem, and yet everything we were doing was classified.” Worse yet, it didn’t have to be that way: “And I knew it didn’t have to be classified, but it was.”1 Decades before the rise of big data in the 1990s and 2000s, the National Security Agency had institutionalized data collection, algorithms, and forms of analysis.
这是怎么发生的？
How did this come to pass?
我们需要搬迁到东伊尔斯利东北六十六英里处，1905 年戈塞特和卡尔皮尔逊曾在那里相遇。那里是英国一个安静的小镇布莱切利园，也是第二次世界大战中最秘密、最重要的地点之一。
We need to relocate to sixty-six miles northeast of East Ilsley, where Gosset and Karl Pearson had met in 1905. There lay the quiet English town of Bletchley Park, one of the most secret and most significant sites in World War II.
布莱切利园
Bletchley Park
当统计学家费舍尔和奈曼为真理和错误而争执不休时，一群局外人正在战争背景下将计算、劳动力和数据结合起来，创造一个截然不同的未来。统计学的局外人是布莱切利园的工程师、语言学家和数学家，他们秘密藏身于英国牛津和剑桥之间，破解德国密码。这些科学家和人文学者以“雷德利上尉的射击队”为掩护，其中许多人是通过老男孩网络招募的，他们实际上是建造专用计算硬件以理解大规模数据流的先驱。2
While the statisticians Fisher and Neyman battled over the truths and errors, a group of outsiders was creating a radically different future combining computation, labor, and data in the context of war. The outsiders to statistics were the engineers, linguists, and mathematicians of Bletchley Park, nestled secretly between Oxford and Cambridge in England, breaking German codes. Protected by a cover story of “Captain Ridley’s Shooting Party,” these scientists and humanists, many recruited through old-boy networks, were in fact pioneers in building specialized computing hardware for making sense of streams of data at a very large scale.2
工作中充斥着机器、纸带、男男女女的嘈杂声，只有少数人在办公桌前工作，他们写的字母比公式还多。大多数在写字的人都来自不同的学术领域，擅长国际象棋或填字游戏等游戏，而不是学术统计学。其中最著名的是艾伦·图灵，他在 1939 年 9 月英国宣战后的第二天乘火车前往布莱切利，他是一位以逻辑学工作而闻名的数学家，之前也从事过一些统计学工作。
The work involved a noisy cacophony of machines, paper tape, men and women, with only a few working at desks scribbling more letters than formulas. Most of the men doing the scribbling were drawn from a variety of academic pursuits and skill in games such as chess or crossword puzzles, not from academic statistics. The most famous of them, Alan Turing, who took a train to Bletchley the day after the United Kingdom declared war in September of 1939, was a mathematician known primarily for work in logic, who had earlier worked some with statistics.
图灵和他的同事并没有专注于量化一个国家人民的素质或研究科学假设，而是致力于一项应用的军事任务，涉及数据和当时世界上最大的计算。
Rather than focusing on quantifying the qualities of a state’s people or investigating scientific hypotheses, Turing and his colleagues worked on an applied, martial task, involving data and the world’s largest (at the time) computation.
这是当时最激进的统计和数据实践，标志着我们更广泛的历史的一个分水岭时刻：数据跃升为由工程和解决问题定义的务实的新存在。
This was the practice of statistics and data in the most aggressive form of its era, and marks a watershed moment in our broader history: when data leapt to a pragmatic new existence defined by engineering and problem-solving.
布莱切利园的研究人员没有参加过关于数理统计学严谨性的激烈争论，因此他们开发了专用计算硬件，并图灵想出了自己的统计方法来破解二战期间使用的“牢不可破”的德国密码（最著名的是恩尼格玛密码机）。事实上，“德国人很清楚恩尼格玛密码机可以被破解的方法。但他们得出结论，需要一整栋楼的设备才能破解它”，据数学家、美国海军上尉、后来的国家安全局研究员霍华德·坎佩恩说。“而我们拥有的就是这样。一整栋楼的设备。” 3破译德国密码需要计算枚举天文数字的假设，每个假设都与当天德国加密机可能的秘密设置相对应。每天，额外的数据都会完善每个假设的概率，初始设置基于对德国军方典型语言的猜测和启发式方法。数学严谨性，即新学术统计学的基石，不是重点。面对生死攸关的任务，图灵和同事使用了现在所谓的“贝叶斯”方法。他们采用了多种方法，使用专用的机电计算设备“炸弹”，这些设备会发出噪音旋转直至“停止”，然后机器会停下来，显示出潜在的解决方案。
Unschooled in the raging debates about rigor in mathematical statistics, the researchers of Bletchley Park developed special-purpose computing hardware along with their own statistical methods for breaking the “unbreakable” German cyphers (most famously the Enigma machine) employed during World War II. In fact, “the Germans were well aware of the way the Enigma could be broken. But they had concluded that it would take a whole building full of equipment to do it” according to mathematician, US naval captain, and later NSA researcher Howard Campaigne. “And that’s what we had. A building full of equipment.”3 Decoding the German codes instead required computational enumeration of an astronomical number of hypotheses, each corresponding to that day’s possible secret settings of the Germans’ encrypting machines. Each day, additional data would refine each hypothesis’s probability, with initial settings based on guesses and heuristics about the typical language employed by the German military. Mathematical rigor, the bedrock of the new academic statistics, was beside the point. Faced with a life-or-death job to be done, Turing and colleagues used what would now be called “Bayesian” methods. They deployed diverse methods using special-purpose electromechanical computing devices, called “bombes,” whirling noisily until a “stop,” when the machines would halt to reveal a potential solution.
对于像图灵这样的天才来说，布莱切利之所以重要，是因为它使数据分析工业化。“1944 年的布莱切利园并不是民间传说中那种小屋式的、学院式的、非正式的组织，”历史学家大卫·肯扬写道。1943 年以后，“任务不是为个人天才提供肥沃的栖息地，而是扩大和工业化密码破译大师开发的技术，并创建系统，让没有牛津剑桥教育水平的员工能够快速地将他们的方法应用于数千个数据项。” 4
For all the work of geniuses like Turing, Bletchley mattered because it made data analysis industrial. “Bletchley Park in 1944 was not the hutted, collegiate, informal organisation of popular myth,” writes the historian David Kenyon. After 1943, the “task was not to provide a fertile habitat for individual genius, but rather to scale up and industrialise the techniques developed by the master codebreakers, and to create systems allowing their methods to be applied to thousands of items of data, at speed, by staff without an Oxbridge level of education.”4
布莱切利的努力最终导致一些历史学家创造了世界上第一台“计算机”。这个词的现代意义：数字、电子和可编程机器，称为巨像。一名工作人员介绍了数据量和机器的性质：“孔眼经过电子眼。它们以每秒五千个的速度通过，因此每秒记录五千个字母。” 5管理这些主要记录在繁琐磁带上的数据需要劳动力，主要是女性的劳动力。布莱切利的工作有明显的性别等级：“所有与巨像一起工作的密码学家都是男性，所有操作员都是女性”，尽管他们中的许多人都接受过大学教育。 6 军方最初要求女性操作员而不是男性数学家进行操练和行军。科学史学家珍妮特·阿巴特 (Janet Abbate) 解释说，这些“对巨像操作员的要求表明了他们的上级不言而喻的假设，即女性的工作本质上是平凡的，不需要一个人的全部精力”。7将数据输入和输出机器是一项艰苦的工作。磁带以每秒四十英尺的速度飞过巨像。一位名叫埃莉诺·爱尔兰的操作员解释说，“这是一项棘手的操作，要让磁带达到正确的张力……[我们] 很担心磁带会断裂。” 8
The Bletchley effort culminated in the creation of what some historians consider the world’s first “computers” in the contemporary sense of the word: digital, electronic, and programmable machines, called the Colossus. One staff member gave a sense of the volume of data and the nature of the machines: “the sprocket holes went past an electronic eye. They went past at five thousand per second, so that five thousand letters were registered per second.”5 Managing this data, registered primarily on finicky tapes, required labor, predominantly the labor of women. Work at Bletchley had a decidedly gendered hierarchy: “all of the cryptographers working with Colossus were men, and all of the operators were women,” even though many of them had received a university education.6 The military initially subjected the women operators but not the male mathematicians to drills and marches. These “demands placed on the Colossus operators reveal their superiors’ unspoken presumption that women’s work is by nature mundane and does not require one’s full energies,” science historian Janet Abbate explains.7 Getting data in and out of the machines was hard work. Tape flew through the Colossus at forty feet per second. One operator, Eleanor Ireland, explained it “was a tricky operation, getting the tape to the right tension . . . [we] were terrified of the tape breaking.”8
战争中时间至关重要。像爱尔兰一样，操作这些机器的女性发誓要保持沉默，直到她们生命的尽头，英国政府才最终解密了布莱切利行动的部分内容。“我最大的悲伤是，”巨人号操作员凯瑟琳·考伊回忆道，“我亲爱的丈夫在 1975 年去世时，不知道我在战争中做了什么。” 9这种保密是如此彻底，以至于信息技术的历史忽视了这些设备和女性团队，历史学家马尔·希克斯解释说，“矛盾的是，这确保了英国的成就在以美国为中心的早期电子计算故事中作为失败者载入史册。” 10
Time mattered in war. And the women like Ireland who staffed the machines pledged silence, which they kept until nearly the end of their lives, when the British government finally declassified aspects of the Bletchley effort. “My great sadness,” reminisced Colossus operator Catherine Caughey, “is that my beloved husband died in 1975 without knowing what I did in the war.”9 So complete was this secrecy that the history of information technologies neglected both the devices and the teams of women, historian Mar Hicks explains, “paradoxically ensuring that British accomplishments went down in history as also-rans in a US-centric story of early electronic computing.”10
与此同时，在美国
Meanwhile, in the States
在大西洋彼岸，美国海军和陆军建造了规模越来越大的工厂级设施，使用新旧机器来处理捕获的轴心国通信，从缩微胶片到 IBM 卡片处理机（称为制表机）。尽管长期保密，但美国和英国逐渐发展出密切的密码学关系。1941 年初，当美国密码学家访问布莱切利时，他们最初并没有被告知用于破解 Enigma 的机器，只是被告知了方法的一般性。11关系迅速升温：到1942年，美国密码学家所罗门·库尔巴克访问了几个月，他说，在访问布莱切利期间，英国人“向我展示了一切，他们对德国系统的操作细节……他们向我展示了炸弹及其操作方式。” 12图灵本人前往美国，参观了制造密码破译炸弹的工厂，并在纽约市的贝尔实验室停留。图灵等人的方法很快就被融入到密码学工作流程中，机器也经过修改，以适应新的工业规模的统计分析。
Across the Atlantic, the US Navy and Army built up increasingly large factory-size installations to process captured Axis communications using machines new and old, ranging from microfilm to IBM card-processing machines called tabulators. Despite long-standing secrecy, the United States and United Kingdom slowly developed a close cryptological relationship. When American cryptologists visited Bletchley early in 1941, they were not initially told about the machines for breaking Enigma—only generalities about the approach.11 Relations warmed quickly: by 1942, American cryptographer Solomon Kullback visited for months, saying that, during his visit to Bletchley, the British “showed me everything, the details of their operations on the German systems. . . . They showed me the bombes and how they operated.”12 Turing himself traveled to the US, visiting the factory making code-breaking bombes as well as stopping at Bell Labs in New York City. The approaches of Turing and company soon were integrated into the cryptological workflow and machines modified to accommodate the new industrial-scale statistical analysis.
“密码学重新安排了权力”，菲利普·罗加韦写道。它“配置了谁能做什么，从什么地方做。” 13第二次世界大战证明了这一点：破解密码决定性地改变了全球权力关系，帮助盟军通过更好的情报在欧洲和太平洋地区取得决定性胜利。1942 年，美国海军密码破译员破解了日本海军的加密通信，使他们能够预测对太平洋中途岛环礁发动袭击的时间和性质。尽管美国海军仍因珍珠港袭击而虚弱不堪，但他们还是能够突袭日本舰队并取得胜利，这为美国重建其耗竭的海军提供了时间。次年 4 月，海军密码分析揭示了日本元帅五十六即将出访的细节山本的英勇行为，使得“复仇行动”得以实施，山本在行动中牺牲，这大大鼓舞了美军的士气，也使日本海军失去了一位被认为是“杰出海军军官”的领导人。14 1944年，盟军在诺曼底登陆时，布莱切利园的分析师和他们的美国同行对德国在法国和低地国家的立场提供了前所未有的了解。15工业数据处理极大地改变了权力关系。战争刚结束，在北约成立前几年，美国和英国巩固了他们在二战中的密码联盟和前所未有的情报共享，建立了一种紧密的密码破译关系，这种关系一直持续到今天——英语国家之间的长期正式联盟这些国家很快就包括加拿大、澳大利亚和新西兰，组成了“五眼联盟”。
“Cryptography rearranges power,” writes Phillip Rogaway. It “configures who can do what, from what.”13 World War II demonstrated this: breaking codes decisively altered worldwide power relations, by helping Allied forces to secure decisive victories through better intelligence in Europe as well as in the Pacific. In 1942, US Navy cryptographers broke encrypted Japanese naval communications that allowed them to predict the timing and nature of an attack on Midway Atoll in the Pacific. While still enfeebled by the attack on Pearl Harbor, the US Navy was able to surprise attack the Japanese fleet and secure a victory that helped give the US time to rebuild its depleted navy. The next April, Navy cryptanalysis revealed details of an upcoming tour by Japanese marshal admiral Isoroku Yamamoto, making possible “Operation Vengeance,” in which he was killed, boosting the morale of US forces and costing the Japanese Navy a leader considered “the outstanding naval officer.”14 When the Allies landed in Normandy in 1944, Bletchley Park analysts and their American counterparts provided an unprecedented level of understanding of the German positions in France and the Low Countries.15 Industrial data-crunching altered power relations dramatically. Just after the war, and years before NATO emerged, the US and the UK cemented their cryptographic alliance and unprecedented intelligence sharing from World War II into a tight code-breaking relationship that persists to this day— a long-standing formal alliance between the Anglophone nations that quickly encompassed Canada, Australia, and New Zealand to make up the “Five Eyes.”
布莱切利园密码破译，1943 年。布莱切利园信托。Getty Images。经确认为由 Dorothy Du Boisson（左）和身份不明的 Wren 操作的 Colossus。识别图来自 Janet Abbate，《重新编码性别：女性在计算领域不断变化的参与度》（麻省理工学院出版社，2012 年），第 15 页。
Code Breaking at Bletchley Park, 1943. Bletchley Park Trust. Getty Images. Identified as Colossus operated by Dorothy Du Boisson (left) and an unidentified Wren. Identification draw from Janet Abbate, Recoding Gender: Women’s Changing Participation in Computing (MIT Press, 2012), p. 15.
这些数据并非用于探寻人类或自然的潜在真相。这些数据并非来自小型实验，记录在小笔记本中。这些数据是出于迫切需要而收集的 — — 即在短时间内提供答案，以推动行动并挽救生命。这些答案只能通过工业规模的数据分析得出。
This was not data in search of latent truths about humanity or nature. This was not data from small experiments, recorded in small notebooks. This was data motivated by a pressing need—to provide answers in short order that could spur action and save lives. Answers that could come only from industrial-scale data analysis.
精明的启发式方法加速了对大量不同假设的计算搜索，并动态更新了由先验信念初始化的概率，这对 Fisher 和 Neyman 来说都是令人厌恶的，但却推动了应用计算统计学的诞生，而应用计算统计学如今已成为企业数据挖掘和人工智能的核心。布莱切利分析的核心是一种被数学家鄙视但在战争期间却受到欢迎并实现工业化的统计形式：贝叶斯统计。
The savvy heuristics expediting the computational search over an astronomical abundance of different hypotheses, and the dynamic updating of probabilities initialized by prior belief, would have been anathema to Fisher and Neyman alike, but set into motion the birth of applied computational statistics now at the heart of corporate data mining and artificial intelligence. Central to analysis at Bletchley was a form of statistics despised by the mathematicians but embraced and made industrial during the war: the Bayesian kind.
贝叶斯：从神到解密
Bayes: From the Deity to Decryption
费舍尔抱怨奈曼和皮尔逊只是数学家，不是科学家；他们反驳说费舍尔只是科学家，不是数学家。然而，在他们这场数学斗争中，最严重的侮辱是被指控为“贝叶斯主义者”。然而，贝叶斯统计方法被证明是一种极好且直接的方法，布莱切利园每天都会用它来评估敌方信息可能的解密结果。
Fisher complained that Neyman and Pearson were mere mathematicians and not scientists; they retorted that Fisher was a mere scientist and no mathematician. The deepest insult of all lobbied in their math battle, however, was the accusation of being “Bayesian.” And yet, Bayesian statistics turned out to be an excellent and straightforward approach for making decisions daily at Bletchley Park in evaluating possible decryptions of the enemy messages.
为了说明贝叶斯方法背后的思想：假设有一所大学正处于 COVID 疫情之中。想象一下，首先，每个学生都接受了“完美”的测试，并且所有阳性和阴性病例都是已知的。然而，一个明显的系统错误导致记录在学生收到通知之前就丢失了。唯一幸存的统计数据是1% 的学生患有这种疾病。学校很快给每个人做了一个快速但不太可靠的测试，其中有 99% 的概率，病人的分数为“阳性”，而健康人的分数为“阴性”的概率为 99%。所有检测呈阳性的学生都被隔离在一个宿舍里。想象一下，你在宿舍里遇到一个学生，你必须决定：这个学生真的生病的概率是多少？如果你是一个贝叶斯主义者，答案就很简单了。
To illustrate the idea behind Bayesian methods: Consider a college in the midst of a COVID outbreak. Imagine that, first, each student has been tested with a “perfect” test, and all positives and negative cases are known. However, a flagrant system error results in the records being lost before students can be informed. The only surviving statistic is that 1 percent of the students have the disease. The college quickly gives everyone a rapid but less reliable test, one in which there’s a 99 percent probability that the score is “positive” for sick people, and 99 percent chance that the score is “negative” for healthy people. All students testing positive are quarantined in one dorm. Imagine you meet a student in the dorm and you must decide: What is the probability this student is actually sick? The answer is straightforward when you’re a Bayesian.
实际上，“贝叶斯学派”仅仅意味着使用贝叶斯规则，即将我们知道的事物与我们想知道的事物联系起来的定义性数学方程。这样的数学规则有什么争议呢？为什么“贝叶斯”令人厌恶？贝叶斯统计不仅指使用方程，还指一种经常被嘲笑为“主观学派”的概率解释。在贝叶斯解释中，某事发生的概率是相信某事会发生的程度。它之所以“主观”，是因为它与人类有关——一个有信念的“主体”。相比之下，直到最近，绝大多数数理统计学家都更喜欢将概率解释为在假设的无数次重复的同一实验中某事发生的客观频率的陈述：例如，一个公平的骰子有六分之一的概率会掷出 5。数学和科学体现了客观性——事实上，主观只是 Fisher 和 Neyman 以不同方式谴责的思维品质。所有这些都至关重要，因为数据最强大的影响力之一就是作为一种援引客观事实的修辞工具。但对贝叶斯的哲学解读（解释问题）与使用被称为“贝叶斯规则”的公式（算术问题）截然不同。
As a practical matter, being a “Bayesian” simply means using Bayes’ rule, a definitional mathematical equation connecting things we know to a thing we want to know. What could be controversial about such a mathematical rule? Why is “Bayes” anathema? Bayesian statistics refers not just to using the equation but to an interpretation of probability often derided as “the subjective school.” Within the Bayesian interpretation, the probability of something occurring is a degree of belief that something will occur. It is “subjective” because it is about a human being—a “subject” having the belief. In contrast, until recently, the vast majority of mathematical statisticians preferred working with probabilities interpreted as a statement of the objective frequency that something will happen in a hypothetical infinite number of repetitions of the same experiment: for example, a fair die will roll a 5 one-sixth of the time. Mathematics and science epitomized objectivity—indeed subjective was just the quality of thinking that Fisher and Neyman deplored in their different ways. All this is crucial, as part of data’s most powerful impact is as a rhetorical tool invoking objective truth. But being philosophical about Bayes (a matter of interpretation) is very different from using the formula known as “Bayes’ rule” (a matter of arithmetic).
让我们回到 COVID 诊断的例子：我们首先要问“一个人感染新冠病毒的概率是多少？假设一个人真的被感染了，那么他检测出 COVID 阳性的可能性有多大？”我们真正想要的是将这句话反过来想：“假设一个人的检测结果为阳性，那么他被感染的概率是多少？ ” 16用英语写成的这句话看起来像是一个小小的文字游戏，但事实上，细微的变化涉及到对我们如何判断什么是真实的以及如何做出决策的不同解释。
Let’s get back to the COVID diagnostic example: We start with “What is the probability that one would get a positive test for COVID, given that one actually is actually infected?” What we really want requires us to turn this around: “What is the probability that one is infected given that one has gotten a positive test?”16 Written in English, this appears to be a small game of wordplay, but in fact, the slight change involves a different accounting of how we decide what is true and how we make decisions.
贝叶斯的令人兴奋的地方在于，我们可以想象通过简单地计算各种概率来做出关于信念的决定。这意味着只需仔细地将每个单独的概率制成表格。问题是什么？问题出在这句陈述中：“众所周知，一百个人中有一个患有这种疾病。”生活中很少会给人这样的已知量（回想一下，在人为的例子中，只有完美的测试与明显的系统错误相结合才能让我们这样做）。只有当我们知道或能够估计某人患有这种疾病的总体概率（无论测试结果如何）时，才能计算出这个概率！
What’s exciting about Bayes is that we can imagine coming to decisions about belief by simply doing accounting of various probabilities. This means merely tabulating carefully each of the separate probabilities. What’s the rub? The problem comes in this statement, “It is known that one in one hundred people has the disease.” Rarely does life hand one such a known quantity (recall that in the contrived example it was only the combination of a perfect test with a flagrant system error that allowed us to do so). This probability can only be calculated if we know or can estimate the overall probability that someone has the disease irrespective of the test result!
这个问题出现在该规则的最初阐述中，即 18 世纪牧师兼学者托马斯·贝叶斯在其死后发表的一篇文章中。这篇文章包含了至关重要的见解，即数据和假设的概率必须是两个项的乘积：给定假设的数据概率，以及假设本身的概率。第二个就是刚才提到的具有挑战性的数量。它通常被称为贝叶斯“先验”：它是一种有趣的东西，因为它“先于”实验数据并且独立于实验数据——并且原则上可以在进行实验之前计算出来。在没有所有实验和观察数据的情况下，知道一个假设是否可能意味着什么？
This problem emerged in the initial articulation of the rule, in a posthumously published essay by one Thomas Bayes, an eighteenth-century minister and scholar. This essay contained the crucial insight that the probability of data and hypothesis must be the product of two terms: the probability of the data given the hypothesis, and the probability of the hypothesis itself. It is the second that is the challenging quantity just mentioned. It’s often called Bayesian “prior”: it’s a funny beast, as it is “prior to” and independent of experimental data—and in principle computable even prior to performing an experiment. What does it mean to know if a hypothesis is probable in the absence of all experimental and observational data?
虽然贝叶斯在原文中没有提到上帝，但历史学家斯蒂芬·斯蒂格勒认为贝叶斯写这篇文章是为了反驳苏格兰哲学家大卫休谟关于基督复活可能性的论证，这是贝叶斯规则的一种运用，一直延续到本世纪。17臭名昭著的怀疑论者休谟希望计算在有奇迹报道的情况下复活的概率。从数学上讲，贝叶斯规则指出，在奇迹被报道的情况下，奇迹为真的概率等于在奇迹为真的情况下被报道的概率乘以存在奇迹的先验概率，再除以奇迹被报道的概率（无论奇迹是否真实）。一个等价但更容易解释的问题是：几率是多少？即：
While Bayes does not mention God in the original essay, the historian Stephen Stigler argues that Bayes wrote it to refute an argument by the Scottish philosopher David Hume regarding the likelihood of Christ’s resurrection, a use of Bayes’ rule which has continued to the present century.17 The notorious skeptic Hume wished to compute the probability of the resurrection given the existence of reports of miracles. Mathematically, Bayes’ rule states that the probability of the miracles being true given that they’ve been reported is the probability that they would be reported given that they were true times the prior probability that there are miracles, divided by the probability that miracles would be reported (whether or not they’re actually true). An equivalent but slightly more interpretable question is: What are the odds? That is:
这里问题变得非常清楚。即使我们对奇迹确实发生时有人报告奇迹的概率达成共识，我们对于奇迹发生的先验概率——即P（有真正的奇迹）——可能存在很大分歧。18如果不对这些先验概率的数值达成一致，我们就无法就奇迹发生的概率达成一致，即使我们对奇迹是否发生都会被报告的可能性达成一致。如此重要的数字对另一个如此主观的数字的依赖一直是数理统计学家长期以来反对贝叶斯立场的关键。例如，一个需要具有神存在的先验概率（就休谟而言）或两个相互竞争的科学假设的先验概率（就奈曼而言）。
Here the problem becomes very clear. Even if we have consensus as to the probability someone would report a miracle given that it really happened, we may have greatly differing opinions as to the prior probability that miracles take place— that is, P(there are real miracles).18 And without agreeing on the numerical value of these prior probabilities, we cannot agree on the probability that miracles took place given that they were so reported, even if we agree on the likelihood they would be reported whether or not they occurred. The reliance of a number so important on another number so subjective has remained the crux of a long-standing anti-Bayesian position among mathematical statisticians. For example, one would need to have such prior probability of divine existence (in the case of Hume) or of two competing scientific hypotheses (in the case considered by Neyman).
尽管存在所有这些严重的反对意见，但正是这种工业规模的贝叶斯分析是艾伦·图灵和布莱切利园密码破译员努力的核心。图灵在布莱切利园撰写的一篇介绍性论文中解释说：“几乎所有概率在密码学中的应用都依赖于因子原理（或贝叶斯定理）。” 19他们在二战期间数字计算的黎明时期，按照贝叶斯规则制定了这种决策。20生病学生案例中的类似方程式是：
Despite all these serious objections, just such Bayesian analysis—at industrial scale—was at the heart of the efforts of Alan Turing and the codebreakers of Bletchley Park. In an introductory treatise he wrote at Bletchley Park, Turing explained, “Nearly all applications of probability to cryptography depend on the factor principle (or Bayes’ theorem).”19 They put just this type of decision, framed in terms of Bayes’ rule, to work at the dawn of digital computation during World War II.20 The analogous equation in the case of the sick students would be:
例如，这应该能为你的决定提供参考，比如，你是否应该呆在那个宿舍外面。
That should inform your decision, for example, as to whether you should stay out of that dorm.
布莱切利的密码破译员不得不依靠启发式和猜测来代替某些先验知识，比如德语字母的使用频率。为什么密码破译员愿意遵循上面列出的记账类型？使用它的一个理由是，在大型数据集的极限下，一个假设的可能性远远超过竞争对手的可能性，以至于决策只依赖于未知的先验知识本身。美国国家安全局的一份文件明确指出指出，“对于密码学家来说，不可能存在任何先验概率分配（无论是否巧妙）会对我们的计算机程序的实用性产生不利影响。” 21当时计算密码学的成功创新现在在数据驱动的应用中很常见。贝叶斯方法非常受欢迎，现在已成为统计复杂性的标志，而不是令人羞愧的东西！22
In place of certain priors, codebreakers at Bletchley had to rely on heuristics and guesses, such as the frequency of letter usage in German. Why were codebreakers willing to follow the types of accounting listed above? One justification for its use is that, in the limit of large data sets, the likelihood for one hypothesis so outweighs that of the competitor that the decision depends only weakly on the unknown priors themselves. An NSA paper explicitly noted, “there can exist for the cryptographer no assignment of a priori odds (whether ingenious or otherwise) that can adversely affect the usefulness of our computer program.”21 What was then a successful innovation for computational cryptography is now commonplace in data-driven applications. With great popularity, Bayesian approaches are now a signifier of statistical sophistication, rather than something to be ashamed of!22
虽然这些技术几十年来一直是机密，但新的计算方法和态度却慢慢地从情报界外传播开来。图灵的工作成果只与少数合作者和盟友分享，但这种方法对大西洋两岸都产生了深远的影响。1942 年，图灵在德国 U 型潜艇巡逻大西洋期间，进行了一次漫长而危险的旅行，前往贝尔实验室与克劳德·香农、约翰·图基和其他未来美国应用计算统计学的杰出人物讨论密码学。图灵的亲密合作伙伴 IJ Good 和 Donald Michie 在接下来的五十年里成为计算统计学和“机器智能”新领域的领导者。战后的几十年里，Good 一直是贝叶斯在统计学中更广泛使用的最执着和最有说服力的宣传者之一。 Good 的大部分职业生涯都在弗吉尼亚理工大学度过，在传播贝叶斯福音的同时，他继续与美国国家安全局及其英国同行政府通信总部 (GCHQ) 密切合作，但合作内容仍处于保密状态。在大量写得很好的论文和书籍中，Good 阐述了使用贝叶斯推理最好构建的有趣统计问题，并经常隐晦地评论说该方法最早是由艾伦·图灵提出的。23在整个冷战期间，美国国家安全局和政府通信总部都在遵循这些发展，开展了丰富的计算统计学项目，但这些项目的方式大多仍是保密的。
While these techniques remained classified secrets for decades, the new computational approaches and attitudes migrated slowly outside of the intelligence world. Turing’s work was shared with only a few collaborators and allies, yet the approach left a deep impact on both sides of the Atlantic. In 1942, Turing made an extended and dangerous trip, given German U-boats patrolling the Atlantic, to Bell Labs to discuss cryptography with Claude Shannon, John Tukey, and other future luminaries of American applied computational statistics. Turing’s close collaborators I. J. Good and Donald Michie went on to become leaders in the new fields of computational statistics and “machine intelligence” over the next fifty years. For decades after the war, Good served as one of the most persistent and convincing proselytizers for the use of Bayes in statistics more generally. Spending much of his career at Virginia Tech, Good continued to collaborate closely with the US National Security Agency and its British counterpart GCHQ (Government Communications Headquarters) in still-classified capacities while spreading the Bayesian gospel. In a torrent of well-written papers and books, Good spelled out interesting statistical problems best framed using Bayesian inference, often with cryptic comments that the method first was proposed by Alan Turing.23 Throughout the Cold War, the NSA and GCHQ pursued a rich program in computational statistics following these developments, in ways that remain mostly classified.
第二次世界大战后
After World War II
直到 20 世纪 80 年代，这种蓬勃发展的实践元素才作为情报界之外的一种亚文化存在，当时计算资源（甚至微型计算机）的可用性使得计算统计学在学术和商业领域得到了广泛的推广。与情报部门内部一样，围绕数学严谨性或概率的正确解释的哲学争论呈现出与学术圈截然不同的特征。
Elements of this thriving practice existed as a subculture outside the world of intelligence until the 1980s, when the availability of computing resources, even on microcomputers, enabled a vast academic and commercial expansion of computational statistics. As within the intelligence services, the philosophical debates around mathematical rigor or the proper interpretation of probabilities took on a character quite unlike that in academic circles.
数学很重要。但数据工程也很重要。数据分析当然需要图灵。但它同样需要工程师和操作员。
Math mattered. But so did engineering for data. Data analysis needed its Turings, to be sure. But it equally needed its engineers, its operators.
对于数据分析而言，与新数学同样重要的是存储和处理数据的工程技术。1948 年有一个悲伤的日子，即所谓的“黑色星期五”，苏联密码突然变得基本无法被美国及其盟国解密。这些密码使 1952 年成立的国家安全局的计算需求转向处理越来越庞大的数据。到 1955 年，超过两千个监听点每月产生 37 吨需要处理的拦截通信，以及 3000 万字的电传通信。仅中国就产生了约 25 万条拦截信息。24美国国家安全局需要处理大量数据的能力，远远超过快速执行算术的能力。数据处理需求非常大，超出了当时的技术水平。历史学家科林·伯克 (Colin Burke) 解释说，在 20 世纪 50 年代中期，“美国国家安全局陷入了美国历史上最大的技术赌博之一：它向计算机公司投入了数千万美元”，以帮助破解苏联的加密系统。25
As important as new mathematics for analyzing data was the engineering to store and to process it. On one sad day in 1948, called “Black Friday,” Soviet encryption abruptly became largely impervious to decryption by the United States and its allies. These codes shifted the computational needs of the National Security Agency, founded in 1952, toward the processing of ever more vast stores of data. By 1955, more than two thousand listening positions produced thirty-seven tons of intercepted communication that needed processing per month, along with 30 million words of teletype communications. China alone produced some 250,000 intercepted messages.24 The NSA needed the capacity to process large amounts of data far more than the capacity to perform arithmetic quickly. The data processing needs were extraordinary and outpaced the technology of the time. Historian Colin Burke explains that in the middle of the 1950s, “NSA became entangled in one of the great techno-gambles in American history: it shunted tens of millions of dollars to computer companies” to help overcome the Soviet encryption system.25
IBM 科学家 Frances Allen 是第一位获得 图灵奖得主描述了 NSA 需要的机器：“一台流媒体机器，它可以从 NSA 在世界各地的监听站收集信息——当时主要是监听俄罗斯和苏联——然后获取大量数据，一些数据是加密的，一些是开放的，并对这些数据进行密码破解。” 26大数据意味着大型机器：“这台机器上连接着一个 [拖拉机] 磁带系统，其中包含大量信息，信息可以从磁带系统流经 Stretch Harvest 内存，通过解码单元、Harvest 单元，然后返回——答案，无论结果如何，都会不间断地返回。” 27她后来解释说：“这是一个巨大的盒式磁带系统，磁带上有地址，并自动编程为拉出磁带，将其带到读取器，然后将其取下，然后读取它。” 28这台机器是一个巨大的模式识别设备，实时处理这些数据流，需要一种为此目的优化的编程语言。29
IBM scientist Frances Allen, the first woman to win the Turing Award, described what NSA needed in a machine: “a streaming machine, which could take information that was gathered from the listening stations that NSA had around the world—mostly listening to Russia at the time, the Soviet Union—and then take that vast amount of data, some coded, some open, and do code breaking on it.”26 Large data meant large machines: “attached to this machine was a [tractor] tape system, which contained vast amounts of information, and the information could stream from the tape system through the Stretch Harvest memory, through the decoding unit, the Harvest unit, and then back out—the answers, whatever the results, back out without ever stopping.”27 She later explained, “It was a great giant cartridge system where the tapes had addresses and were automatically programmed to pull up a tape, bring it up to a reader and then take it off, then read it.”28 The machine was a giant pattern recognition device working in real time on this data stream and needed a programming language optimized for this purpose.29
在“关注大量数据处理以及非数值逻辑处理的巨大灵活性和多样性”方面，美国国家安全局的需求更类似于大企业，而非物理学家。30就像大量联邦资金推动了更快的算术机器的创造一样，大量联邦密码学资金也资助了更大规模存储机制的大量研究。在20世纪 50 年代中期，IBM 试图实现能力飞跃，而这两者在资金上产生了巨大摩擦。
In focusing “on the manipulation of large volumes of data and great flexibility and variety in non-numerical logical processes,” the NSA had needs more akin to large businesses than to physicists.30 Just as substantial federal funds promoted the creation of ever faster arithmetical machines, substantial federal funds for cryptography sponsored intense work on larger storage mechanisms. The two came together, with great friction, in funding IBM’s attempts to create a jump in capability in the mid-1950s.
一位 NSA 先驱曾沉思道，如果密码学家制造了第一台计算机，那么它们的名字可能会是“分析器”或“信息处理器”，甚至是“数据分析器”。31在美国核武器国家实验室的赞助下，计算机的发展在很大程度上集中于提高模拟所需的处理速度。爆炸。他们需要大量的乘法，而不是大规模的数据分析。*
If cryptologists had made the very first computers, one NSA pioneer mused, their name might have been “analyzers” or “information handlers” or, even, “datalyzers.”31 Under the sponsorship of the US national laboratories concerned with nuclear weapons, computer developments focused to a great extent upon improving the processing speed needed for simulating explosions. They needed lots and lots of multiplication, not large-scale data analysis.*
美国国家安全局曾为 IBM 和雷明顿提供资金，就像他们后来大力资助控制数据公司 (CDC) 和克雷一样，以制造能够更快执行算术的计算机，但或许更重要的是，它们能够处理更多的数据，通常是并行和实时的。32从 20 世纪70年代起，美国国家安全局失去了对超级计算机未来设计的大部分控制权，但它仍然是此类机器的主要市场（如果不是唯一的话）。
The NSA funded IBM and Remington, just as they would later heavily fund Control Data Corporation (CDC) and Cray, to create computers that could perform arithmetic faster, but—perhaps even more importantly—contend with more data, often in parallel and in real time.32 From the 1970s onward, NSA would lose much of its control over the future design of supercomputers, but it remains a primary— if not the primary—market for such machines.
数据即工程
Data as Engineering
美国国家安全局内部的社区以统计学的方式处理这些数据，就像处理布莱切利园的数据一样：将其视为工程问题而非科学问题。†
Communities within the NSA approached the data with statistics much like the figures at Bletchley Park: as an engineering problem more than a scientific one.†
他们需要不同的计算机。他们需要不同的数学，即布莱切利园传统上的数学。尽管美国国家安全局的数学仍处于高度机密状态，但少数解密的文献表明，该机构追求的不仅仅是计算统计，而是实时接收数据的大规模计算统计。美国国家安全局的人员拥有极高的数学水平，拥有不断增长的数据流和专门为处理数据而定制的计算机；然而，与学术统计学家不同，他们不必努力证明自己是数学家。效率是关键：大规模计算的成本在解密的美国国家安全局文件中占据着中心位置，即使所有有趣的部分都被删减了。在一篇关于判断大量假设是否正确解密消息的贝叶斯论文中，分析产生了分析所需的昂贵估计量：“这几乎与实际测试假设一样昂贵。因此，从通信安全的角度来看，上面的表达式”对它来说“没有实际用处”。33 通信安全的观点不需要学术界所要求的纯粹性，而是在统计严谨性和大量数据要求之间取得平衡。另一篇论文指出：“在密码分析中，我们经常进行一百万次或更多次连续实验，并为每个实验计算一个贝叶斯因子。”事实上，该期刊上的论文明确拒绝了统计学家和哲学家在没有先验概率的情况下使用贝叶斯的担忧。34鉴于任务，除了我们在论文中看到的哲学和统计价值观之外，还有其他价值观Fisher 和 Neyman 必将获胜。贝叶斯分析在大规模情况下过于强大。35
They needed different computers. And they needed different math, math in the tradition of the work at Bletchley Park. While NSA mathematics remains highly classified, a small number of declassified works show that the agency pursued not simply computational statistics, but large-scale computational statistics on data being received in real time. The NSA had personnel with a tremendous degree of mathematical sophistication combined with ever-increasing streams of data and computers custom-built to contend with data; unlike academic statisticians, however, they did not have to work to justify themselves as mathematicians. Efficiency is key: the cost of large-scale computation figures centrally in the declassified NSA papers, even with all the juicy bits redacted. In a Bayesian paper on judging large numbers of hypotheses about the proper decryption of a message, the analysis produced a costly estimator needed in the analysis: “This would cost almost as much as doing the actual testing of the hypotheses. Hence, from a COMSEC [communications security] point of view, the above expression” for it “is not practically useful.”33 A communications security point of view requires not the purity demanded by academics, but a balance between statistical rigor and the requirements of vast data. “In cryptanalysis,” another paper notes, “we frequently perform a million or more consecutive experiments, with a Bayes Factor computed for each experiment.” Indeed, papers in the journal explicitly reject the concerns of statisticians and philosophers with the use of Bayes in the absence of a priori probabilities.34 Given the mission, values other than the philosophical and statistical ones like we saw in Fisher and Neyman must prevail. Bayesian analysis was too powerful at great scale.35
在 20 世纪 90 年代以来，利用日常商业交易自动积累的数据的大规模算法模型颠覆了媒体和广告业之前，美国国家安全局内部已经开发了一种计算密集型统计机器学习，专注于实时生成的大量杂乱数据。与未来的机器学习一样，它大量但有选择地借鉴了统计数据，与当代数据分析一样，该机构也在努力满足实际数据库的需求，但目的不同。
Before large-scale algorithmic models drawing upon the automatically accumulated data of everyday business transactions upended media and advertising from the 1990s onward, the NSA had internally developed its form of computational heavy statistical machine learning focused on high volumes of messy data generated in real time. Like future machine learning it drew heavily but selectively upon statistics, and like contemporary data analysis the agency wrestled with the demands of practical databases, but with different ends.
然而，所有这些工作都是保密的，对数据的计算态度和存储技术的转换逐渐进入了非机密世界，也许最著名的是以两位美国国家安全局科学家命名的统计距离，即库尔贝克-莱布勒散度。
However classified all this work, computational attitudes toward data and transformations in storage technologies slowly found their way into the nonclassified world, perhaps most famously a statistical distance named after two NSA scientists, Kullback-Leibler divergence.
科学界也紧随其后。1950 年，海军研究办公室的米娜·里斯 (Mina Rees) 指出，早期的机器“非常重视”“它们可以接受少量信息，对这些信息进行非常快速的广泛操作，并得出少量信息作为答案”。她写道，现在的兴趣“似乎在于进一步探索使用机器接受大量数据，对它们进行非常简单的操作，并打印出可能非常大量的结果”。36高能物理产生的实验数据迅速挑战了存储和处理能力。37在科学和侦查中，可能被分析和存储的数据已经超过了处理能力、内存和存储容量。2009 年《科学》杂志的一篇文章指出：“在过去 40 多年里，摩尔定律使硅片上的晶体管变得更小，处理器变得更快。与此同时，存储磁盘的技术改进无法跟上计算机速度越来越快所生成的大量科学数据。” 38
The sciences followed suit. In 1950, Mina Rees of the Naval Research Office noted the “great emphasis” on early machines “that would accept a small amount of information, perform very rapidly extensive operations on this information, and turn out a small amount of information as its answer.” Now, she wrote, the interest “seems to lie in a further exploration of the use of machines to accept large amounts of data, perform very simple operations upon them, and print out, possibly, very large numbers of results.”36 The experimental data produced in high-energy physics quickly challenged storage and processing abilities alike.37 In science as in snooping, the data potentially to be analyzed and stored has ever outstripped processing power, memory, and storage capacity. “Over the past 40 years or more,” a piece in Science noted in 2009, “Moore’s Law has enabled transistors on silicon chips to get smaller and processors to get faster. At the same time, technology improvements for disks for storage cannot keep up with the ever-increasing flood of scientific data generated by the faster computers.”38
第二次世界大战后，这种思想脉络或许没有哪个地方像美国电话电报公司 (AT&T) 的贝尔实验室那样蓬勃发展，那里的数据不是关于代码和密码，而是关于更普遍的通信：美国境内和境外的电话通话。
Perhaps nowhere did this intellectual thread flourish in the wake of World War II as at AT&T’s Bell Labs, where the data was not about codes and ciphers but about communications more generally: phone calls across the United States and abroad.
实验室数据
Data at the Labs
我们（美国国家安全局）与贝尔实验室保持着非常密切的联系。
We [NSA] had very close contacts with the Bell Laboratories.
可以说，他们非常愿意与我们合作。
They were very, let’s say, willing to work along with us.
— 所罗门·库尔巴克 (Solomon Kullback，1907-1994)，1942 年在布莱切利工作，之后担任美国国家安全局首席科学家，职业生涯十分出色，1982 年接受采访时表示
—Solomon Kullback (1907–1994), who spent 1942 at Bletchley before a distinguished career as chief scientist at NSA, interviewed in 1982
和布莱切利和美国国家安全局一样，贝尔实验室也是利用数据进行计算的早期典范。贝尔的数据主要关注人及其通信——这比通信成为互联网命脉早了几十年。
Like Bletchley and the NSA, Bell Labs was an early example of computing with data. And Bell’s data focused on people and their communications—decades before those became the lifeblood of the internet.
AT&T 的贝尔实验室是当时的 Google Research，它直接处理人们的数据和信息，政府垄断了所有的数据、所有的研究人员和所有的计算能力。虽然贝尔研究人员与学术界保持着密切联系，但他们强调自己的工作与学术传统和口号截然不同。
The Google Research of its day, AT&T’s Bell Labs worked directly with data and information about people within a government-tolerated monopoly with all the data, all the researchers, and all the computing power. While they maintained close ties with academia, Bell researchers emphasized their work’s distinctness from the academic traditions and shibboleths.
1962 年，普林斯顿贝尔实验室数学家约翰·图基在一份宣言中呼吁采用一种新方法，他称之为“数据分析”，这种方法更专注于发现，而不是通过数学证明进行确认。作为科学实际上，图基认为，数据分析是一门艺术，一种判断形式，而不是一门逻辑上封闭的学科，他鼓励创造新的工具，从方格纸到计算机图形学，以实现发现。
In a 1962 manifesto, the Princeton-Bell Labs mathematician John Tukey called for a new approach he dubbed “data analysis” that would be more dedicated to discovery than to confirmation through mathematical proof. As a scientific practice, Tukey argued, data analysis is an art, a form of judgment, not a logically closed discipline, and he encouraged the creation of new tools, from graph paper to computer graphics, to enable discovery.
斯皮斯率先进行了大规模数据存储，因为第二次世界大战后不久，美国情报界就意识到了数据存储的必要性。商界很快开始迎头赶上。从 20 世纪 60 年代的航空预订系统数据开始，行业开始以极快的速度积累有关客户的数据。在随后的二十年里，企业收集了日常交易的数据：在特定地点使用信用卡购物、航空旅行、租车，以及后来在图书馆结账。在数十年的商业计算机发展过程中，IBM 等多家公司也采用了计算机，寻求将数据（主要是消费者数据）转化为利润的新方法。到 20 世纪 70 年代中期，越来越多的自由主义者、政府官员和消费者安全发言人注意到了这一点。洛克菲勒基金会负责人指出：“我们开始认识到，有组织的知识将巨大的权力交到了那些不辞辛劳掌握它的人手中。” 39
Spies pioneered large-scale data storage, as its necessity became apparent within the American intelligence community soon after the Second World War. The business world soon began catching up. Starting with the data from airline reservations systems in the 1960s, industry began accumulating data about customers at a rapidly accelerating rate. In the subsequent twenty years, corporations collected data of everyday transactions: credit card purchases at particular locations, airline trips, car rentals, and later, checkouts at libraries. Over the decades of development of computers for business purposes, they were adopted by a variety of other companies such as IBM who sought new ways to turn data—primarily data about consumers—into profit. By the mid-1970s, a growing number of libertarians, government officials, and spokespersons for consumer safety took notice. “We are coming to recognize,” the head of the Rockefeller Foundation noted, “that organized knowledge puts an immense amount of power in the hands of people who take the trouble to master it.”39
尽管战时人们曾体验过数据及其威力，此后数据计算在工业上蓬勃发展，但 20 世纪 40 年代学者和数学家对新型数字计算机的期望主要集中在将其作为逻辑机器而非数据处理器。正如统计学家倾向于抽象数学一样，自 1950 年以来爆发式增长的智能机器的早期支持者大多专注于逻辑和数学，而不是有关人和事物的数据。
Despite the wartime experience with data and its powers, and the industrial flourishing of computing with data thereafter, hopes for the new digital computers among academics and mathematicians in the 1940s focused on them as logical machines—not data processors. Just as statisticians gravitated toward abstract math, most of the early proponents of intelligent machines that exploded from 1950 onward focused on logic and math, not data about people and things.
*原子能委员会的“计算机要求强调高速乘法，而美国国家安全局的重点则是处理大量数据以及非数值逻辑处理的巨大灵活性和多样性。” Samuel S. Snyder，《密码组织推动的计算机进步》，《计算机历史年鉴》 2，第 1 期（1980 年）：66。
* The Atomic Energy Commission’s “computer requirement emphasized high- speed multiplication, whereas the NSA’s emphasis was on manipulation of large volumes of data and great flexibility and variety in non- numerical logical processes.” Samuel S. Snyder, “Computer Advances Pioneered by Cryptologic Organizations,” Annals of the History of Computing 2, no. 1 (1980): 66.
† “我们今天发现自己开始成为一家工厂。对某些人来说，当你看不到数据时，就没那么有趣了。我认为最大的进步之一是目标国家开始使用电传打字机设备并开始以电子方式发送数据。我们曾经以为我们会有一英里半的卡片，整个大楼里都会挤满打卡机操作员来打卡所有数据；但幸运的是，目标国家开始成为我们的主要打卡机操作员，这使我们能够以电子方式转发这些数据。我们目前每天通过电路处理一些 [已编辑] 直接进入大楼并自动处理的数据。……这些数据中的大部分从未被任何特定的人看到过。在某些情况下，结果会在不到一分钟的时间内返回，实际上从未被个人看到过。这并不意味着在数据准备过程中不需要进行大量的分析工作。” Joseph Eachus 等人，《在 NSA 与计算机一起成长（绝密的 Umbra）》， NSA 技术期刊特刊（1972 年）：14。
† “We find ourselves today in the position of beginning to be a factory. To some it is not as much fun when you don’t see the data. I think one of the biggest developments was when the target countries began to use teletype equipment and began to send their data electrically. We thought at one time that we would have a mile and a half of cards, and that we would have the whole building filled with key punch operators to punch all the data; but fortunately the target countries began to be our key punch operators, which led to our being able to forward this data electrically. We are currently handling by electrical circuits some [redacted] per day which come directly into the building and are handled automatically. . . . Much of this data is never seen by any particular person. In some cases, the results go back within less than a minute, having really never been seen by an individual. That doesn’t mean that much analytic work doesn’t go into the preparation of the data.” Joseph Eachus et al., “Growing Up with Computers at NSA (Top Secret Umbra),” NSA Technical Journal Special Issue (1972): 14.
1952 年，贝尔实验室的科学家克劳德·香农在写给一位前教师的信中写道：“我最大的梦想是，有一天能制造出一台真正能够思考、学习、与人类交流并以相当复杂的方式操纵环境的机器。” 1第二次世界大战结束后，工程师、数学家、社会学家和神经学家都在猜测：机器能否执行以前被视为专属于人类智能的任务？一个关键问题是，谁的智能？数学家？语言学家？计算器？还是面包师？在世界上所有的智能形式中，第二次世界大战后几年里，人们最常给出的答案是研究人员等人优先考虑的那种智能：证明定理、下棋、有效地驾驭官僚体系。
“My fondest dream,” the Bell Labs scientist Claude Shannon wrote a former teacher in 1952, “is to someday build a machine that really thinks, learns, communicates with humans and manipulates its environment in a fairly sophisticated way.”1 In the wake of World War II, engineers, mathematicians, sociologists, and neurologists all speculated: Might machines perform tasks previously viewed as exclusively the province of human intelligence? A key question was, Whose intelligence? Mathematicians? Linguists? Calculators? Expert bakers? Of all the forms of intelligence in the world, the answer most typically given in the years after World War II was intelligence of the sort that people like the researchers prioritized: proving theorems, playing chess, navigating bureaucratic systems efficiently.
你可能认为数据分析是这个项目的核心。但事实并非如此。如今，人工智能主要意味着对海量数据集进行机器学习。但当时并非如此。2
You might expect the analysis of data to be central to this project. It wasn’t. Today, artificial intelligence primarily means machine learning on huge data sets. It didn’t then.2
图灵
Turing
1950 年，艾伦·图灵发表了一篇划时代的论文，为机器执行一系列通常被认为需要智能的活动的可能性辩护。他驳斥了计算机不可能具有原创性、它们只能遵循规则而不能适应、它们无法从现实世界中学习的论点。图灵以其在逻辑方面的成就而闻名，但他并没有过分推崇逻辑是人类智能的巅峰。他的观点更加普世，涵盖了广泛的创造性、智能甚至情感活动。
In 1950 Alan Turing published an epochal paper defending the possibility of machinery performing a range of activities typically thought to require intelligence. He rebutted arguments that computers couldn’t be original, that they could only follow rules without adapting, that they couldn’t learn from experience in the world. Famed for his results in logic, Turing was not unduly celebratory of logic as the pinnacle of human intelligence. His views were far more ecumenical, covering a wide range of creative, intelligent, even emotional activity.
在布莱切利工作之前，艾伦·图灵在数学和逻辑史上发表了一项重要成果，比任何数字计算机出现都要早几年。他提出了抽象通用机器（现在称为“图灵机”）的概念，这种机器几乎可以执行任何逻辑运算。战争期间，他和布莱切利园的其他人知道如何从大量数据中得出初步结论。他们晚上都在推测机器智能行动的可能性。战争结束后，图灵设想了各种各样的机器，它们可能能够利用逻辑和数据做出看似智能的行为。3
Before his time in Bletchley, Alan Turing had published a critical result in the history of mathematics and logic, several years before any digital computer. He introduced the idea of an abstract universal machine (now referred to as a “Turing machine”) that could perform nearly any logical operation. During the war, he and others at Bletchley Park knew as much as anyone about drawing tentative conclusions from masses of data. They spent their evenings speculating on the possibility of machines acting intelligently. After the war, Turing envisioned a variety of machines that might be capable of doing apparently intelligent acts drawing upon logic and data alike.3
在他的论文《计算机器与智能》中，图灵将一个猜测隐藏的人是男人还是女人的客厅游戏转化为一种辨别机器是否表现出智能行为的操作方法，即模仿游戏：“对于询问者来说，游戏的目标是确定两个隐藏的人（分别称为 A 和 B）中哪个是男人，哪个是女人。”图灵建议用机器代替男人 A：
In his paper “Computing Machinery and Intelligence,” Turing converted a parlor game of guessing whether a hidden person was a man or woman into an operational approach to discerning whether a machine exhibited intelligent behavior, the imitation game: “The object of the game for the interrogator is to determine which” of two hidden people called A and B “is the man and which is the woman.” Turing suggests replacing the Man A with a machine:
我们现在要问的问题是：“如果一台机器在这场游戏中扮演 A 的角色，会发生什么？”当机器扮演 A 的角色时，询问者做出错误决定的频率会不会和机器扮演 A 的角色一样高？游戏怎么会这样玩，就像男人和女人之间玩游戏时他那样玩？这些问题取代了我们最初的问题“机器能思考吗？”
We now ask the question, “What will happen when a machine takes the part of A in this game?” Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman? These questions replace our original, “Can machines think?”
图灵不是通过询问机器是否像人类一样思考来判断机器是否在思考，而是要求我们检查机器的行为。“如果……一台机器能够被制造出来并令人满意地玩模仿游戏，我们就不必为机器不像人类那样运作而烦恼，因此不能被认为是在有意义地思考。”
Rather than gauging whether a machine is thinking by asking if it thinks as human beings do, Turing asks instead for us to examine its behavior. “If . . . a machine can be constructed to play the imitation game satisfactorily, we need not be troubled” by the objection that machines don’t operate in the ways that human beings do, and thus can’t be thought of as thinking in a meaningful sense.
在他的论文中，图灵对机器智能进行了广泛的思考。尽管他是一位杰出的逻辑学家，但他将经验和数据置于核心地位，并没有只考虑数学之类的形式推理或象棋之类的游戏。他甚至还包括了通常不被认为是机器的活动：
In his paper, Turing mused expansively about machine intelligence. Despite his prominence as a logician, he gave experience and data a central role, and did not consider only formal reasoning like mathematics or games like chess. He even included activities typically not thought of as machinelike:
善良、足智多谋、美丽、友好、有主动性、有幽默感、明辨是非、犯错误、坠入爱河、享受草莓和奶油、让别人爱上它、从经验中学习、恰当地使用语言、成为自己思想的主体、拥有与男人一样多样化的行为、做一些真正新的事情。
Be kind, resourceful, beautiful, friendly, have initiative, have a sense of humour, tell right from wrong, make mistakes, fall in love, enjoy strawberries and cream, make someone fall in love with it, learn from experience, use words properly, be the subject of its own thought, have as much diversity of behaviour as a man, do something really new.
他认为，我们之所以怀疑机器能够做到所有这些事情，是因为我们对有限、丑陋的机器的体验，这些机器都是为“特殊目的”而制造的。从这种有限的经验中，我们错误地得出结论，认为机器无法做到这些事情。他认为，真正限制现有计算机的是计算机内存，他称之为存储——“大多数机器的存储容量非常小”。4大内存计算机可以表现出许多不同的行为：“有人批评机器不能有多种行为，其实就等于说机器不能有太大的存储容量。” 5而产生所有这些结果的基础是机器自我修改的能力：“机器无疑可以成为它自己的主题。它可以用来帮助编写自己的程序，或者预测改变自身结构的影响。通过观察自身行为的结果，它可以修改自己的程序，以更有效地实现某些目的。这些都是不久的将来的可能性，而不是乌托邦式的梦想。” 6图灵在英国政府对他实施化学阉割后英年早逝，人们无法得知他广阔的视野可能带来什么。
Our incredulity that machines might do all these things rests, he argued, on our experience of limited, ugly machines, all made for a “special purpose.” From this limited experience, we mistakenly conclude that machines could do none of these things. The real thing limiting existing computers, he argues, is computer memory, which he calls storage—“the very small storage capacity of most machines.”4 A large memory would enable computers to exhibit many different behaviors: “The criticism that a machine cannot have much diversity of behaviour is just a way of saying that it cannot have much storage capacity.”5 And fundamental to producing all these results would be the ability of machines to modify themselves: “a machine undoubtedly can be its own subject matter. It may be used to help in making up its own programmes, or to predict the effect of alterations in its own structure. By observing the results of its own behaviour it can modify its own programmes so as to achieve some purpose more effectively. These are possibilities of the near future, rather than Utopian dreams.”6 Turing’s untimely death by his own hand in the wake of his chemical castration by the British state foreclosed finding out where his capacious vision might have led.
在他的智能机器愿景中，图灵将从大内存中数据中学习与计算机自我重新编程结合在一起。这是令人头疼的事情。人类学家露西·萨奇曼认为，人工智能方面的努力“是揭示人类假设的强大工具”。图灵揭示了一个广阔的智能愿景，它取材于人类和动物世界，充满了逻辑、爱、创造力、工艺和笑声。在随后的几年里，许多追求机器智能形式的人大大缩小了视野。让机械设备模仿人类智能首先在人类允许自己像机器一样行事的领域取得成功，例如使用算法规则或玩简单的基于规则的游戏。在此过程中，所讨论的智能概念本身失去了图灵所暗示的大部分容量。与此同时，数据和经验失去了创造智能行为的核心地位。如何——以及为什么？
In his vision of intelligent machines, Turing wedded learning from data in large memory stores to computers reprogramming themselves. This was heady stuff. The anthropologist Lucy Suchman argues that efforts in artificial intelligence function “as a powerful disclosing agent for assumptions about the human.”7 Turing disclosed a capacious vision of intelligence, drawn from the human and animal world, full of logic, love, creativity, craft, laughter. In the years to follow, many pursuing forms of machine intelligence narrowed their sights considerably. Getting mechanical devices to imitate human intelligence first succeeded in arenas when humans have allowed themselves to behave like machines, as with algorithmic rules of production or in playing simple rule-based games. Along the way, the very notion of intelligence at issue lost much of the capaciousness Turing suggested. And in parallel, data and experience lost their centrality for the creation of intelligent behavior. How—and why?
第二次世界大战后出现的新计算机将数值计算、信息处理和根据逻辑规则的符号操作结合在一起。 原子弹制造者推崇计算，工业和密码学家推崇数据处理，还有一些人则更注重逻辑。战后机器智能的一个关键派别坚持认为，人类智能最显著的特征是逻辑和符号思维，而不是基于感官经验（数据）或进行大量计算的低级能力。20 世纪 50 年代中期，逻辑派最狂热的拥护者——一位名叫约翰·麦卡锡的年轻数学家——担心专注于数据的研究人员拥有太多影响力。他解释说，使用数据不会产生智能行为：“直接将试错法应用于感官数据和运动活动之间的关系不会导致任何非常复杂的行为。”只有通过抽象感官数据，才能出现更复杂的行为。8需要采取一些措施让机器智能回到正轨。
The new computers emerging in the wake of World War II combined numerical calculation, information processing, and the manipulation of symbols according to logical rules. Atomic bomb makers celebrated calculation; industry and cryptographers celebrated data processing; and still others focused rather on logic. A key faction in postwar machine intelligence insisted that human intelligence was most characterized by logical, symbolic thinking, and not in the lower capacities of working from sense experience (data) or in performing abundant calculations. In the mid-1950s, the most avid partisan of the logical side—a young mathematician named John McCarthy—was concerned that researchers focused on data were holding too much sway. He explained that using data would not create intelligent behavior: “the direct application of trial-and-error methods to the relation between sensory data and motor activity will not lead to any very complicated behavior.” More complicated behavior would emerge only by abstracting away from sensory data.8 Something needed to be done to get machine intelligence back on the right path.
等一下，我们听到了你的呼声。科学家为什么要反对数据？如果战时科学活动的一方面导致了一个秘密数据密集型国家，那么另一方面则导致计算机可以模拟人类智能的想法盛行，这种智能被更狭义地理解为以编程到计算机中的规则表达的符号推理，而不是从数据中推断。
Wait a second, we hear you cry. Why would a scientist be against data? If one side of wartime scientific activity led to a secret data-intensive state, another led to the flourishing of ideas that computers might emulate human intelligence understood more narrowly as symbolic reasoning couched in rules programmed into computers, not as inferences from data.
反对数据：没有测量的数学
Against Data: Mathematics without Measurement
基于大量数据的应用统计学是第二次世界大战的核心。矛盾的是，在战争结束后，许多科学家的心被将社会科学变成更像抽象的纯数学的学科的愿景所吸引，而不是被利用数据理解社会的愿景所吸引。科学史学家阿尔玛 斯坦加特解释说：“战后社会科学数学化的特征不是测量和量化，而是公理化。” 9例如，伟大的法国人类学家克劳德·列维-斯特劳斯在反思社会科学的前沿时，于 1954 年指出，研究人类的人需要摆脱量化；他们需要放弃数据的积累，转而采用抽象的数学和逻辑处理。关于人类的新数学，他写道：“它所关注的领域不是通过大量数据的积累揭示的无穷小变化。”事实上，对人类的研究应该“坚决摆脱‘大数字’的绝望——社会科学迷失在数字的海洋中，无助地依附于这个木筏。” 10列维-斯特劳斯抱怨说，社会科学“仅仅借用了量化方法……这些方法被认为是传统的、在很大程度上已经过时了。”新的“定性数学”表明“严格的处理不再必然意味着诉诸测量。” 11
Applied statistics based on large amounts of data was central to the fight in World War II. Paradoxically, in the wake of the war, the hearts and minds of many scientists were won by visions of making the social sciences into subjects more like abstract pure mathematics rather than by visions of understanding society using data. Science historian Alma Steingart explains, “It was not measurement and quantification that characterized the mathematization of the social sciences after the war, but axiomatization.”9 Reflecting on the cutting edge of social sciences, the great French anthropologist Claude Lévi-Strauss, for example, argued in 1954 that those studying human beings needed to escape quantification; they needed to set aside the amassing of data in favor of an abstract mathematical and logical treatment. Of the new mathematics of human beings, he wrote that “[t]he field with which it is concerned is not that of the infinitesimal variations revealed by the accumulations of vast accumulations of data.” In fact, the study of human beings ought to be “resolutely determined to break away from the hopelessness of the ‘great numbers’—the raft to which the social sciences, lost in an ocean of figures, have been helplessly clinging.”10 Lévi-Strauss complained that the social sciences have “simply borrowed quantitative methods which . . . are regarded as traditional and largely outmoded.” The new “qualitative mathematics” shows “a rigorous treatment no longer necessarily means recourse to measurement.”11
在前面的章节中，我们看到统计学从关注数据收集转向创建数学模型；同样，在第二次世界大战之后，社会学、经济学和政治学等寻求成为科学的学科也从主要关注从经验数据进行概括转向寻求更普遍、更简单、更抽象的理论。在第二次世界大战之后，数学和逻辑理论（关于人类决策、经济、智力）受到重视和推崇。积累数据虽然非常重要，但与广义理论（尤其是以抽象数学术语呈现的理论）相比，它显得黯然失色。
In an earlier chapter we saw how statistics moved away from its focus on the collection of data to the creation of mathematical models; so too, in the wake of World War II, did enterprises seeking to be sciences, such as sociology, economics, and political science, turn away from a dominant focus on generalizing from empirical data toward the seeking of more general, simplifying, abstract theories. In the wake of World War II, mathematical and logical theories—of human decision-making, of the economy, of intelligence— were prized and celebrated. Accumulating data, for all its importance, paled next to generalized theories, particularly those presented in abstract mathematical terms.
思考太重要了，无法量化。在很多领域，研究人员主张人类是理性的假设制定者，受政策的约束，而不是受数据驱动。这些争论还涉及到科学最典型的特征。它们涉及到人类最独特的特征。这些不同的观点对什么是真正的科学提出了截然不同的看法。以及人类是什么。
Thinking was far too important to be reduced to quantification. Across a swath of fields, researchers advocated the idea that humans were rational hypothesis formers, programmed by policies, not driven by data. These debates also concerned what most characterized science. And they concerned what was most distinctive about human beings. These different views offered radically different ideas of what real science is. And of what humans are.
基于规则或符号的人工智能正是在这些反统计的海洋中畅游。理解语言或思维并不需要积累大量数据——事实上，这些数据可能会成为阻碍。理解和模仿人类智能需要抽象和“模式”。它需要公理和规则。它不需要数据驱动的算法。
Rules-based or symbolic artificial intelligence swam in just these anti-statistical seas. Understanding language or thinking did not require the accumulation of vast data—in fact such data would probably get in the way. Understanding— and emulating—human intelligence required abstraction and “schemas.” It needed axioms and rules. It didn’t need data-driven algorithms.
我们将会看到，计算统计和数据并没有消失。但从数据中推断显然不是人工智能在最初几十年的目标。这种反统计倾向在近半个世纪的时间里一直是人工智能的特征。1984 年对人工智能的定义解释说，该领域涉及“符号、非算法的问题解决方法”，因为“人们对医学等学科的大部分知识不是数学或定量的。”这些方法不是“数学或数据处理程序”，而是“定性推理技术”和“理论定律和定义”。12换句话说，是规则，而不是数据。
As we will see, computational statistics and data did not disappear. But inference from data decidedly wasn’t the goal of what came to be called AI in its first decades. This anti-statistical bent characterized AI for almost half a century. A definition of AI from 1984 explained that the field deals “with symbolic, nonalgorithmic methods of problem solving,” as “most of person’s knowledge of a subject like medicine is not mathematical or quantitative.” Rather than “mathematical or data-processing procedures,” the methods involve “qualitative reasoning techniques,” and “theoretical laws and definitions.”12 In other words, rules, not data.
打造“人工智能”
Confecting “Artificial Intelligence”
数学家约翰·麦卡锡是符号方法的热情倡导者，他经常被认为是“人工智能”一词的发明者，包括他自己：“我发明了人工智能这个术语，”他解释说，“当时我们正试图为暑期研究筹集资金”，以“实现达到人类水平智能的长期目标”。这项“夏季研究”名为“达特茅斯人工智能夏季研究项目”，申请的资金来自洛克菲勒基金会。当时麦卡锡是达特茅斯学院的一名初级数学教授，他的前导师克劳德·香农曾帮助他向洛克菲勒基金会推销人工智能。麦卡锡这样描述这个术语的定位：“香农认为人工智能这个术语太过浮夸，可能会引起不利的关注。”然而，麦卡锡希望避免与现有的“自动机研究”（包括“神经网络”和图灵机）领域重叠，并宣布开辟一个新的领域。“所以我决定不再打假旗了。” 13这个野心是巨大的；1955 年的提案声称“学习的每个方面或智能的任何其他特征在原则上都可以得到如此精确的描述，以至于可以制造出机器来模拟它。” 14麦卡锡在 1956 年的会议上，最终得到的是更多的大脑建模者，而不是他想要的公理数学家，这次会议后来被称为达特茅斯研讨会。15这次活动汇集了各种各样、往往相互矛盾的努力，旨在让数字计算机执行被认为是智能的任务，但正如人工智能历史学家乔尼·佩恩所说，研讨会上缺乏心理学专业知识，这意味着对智能的解释“主要由一组在人文科学之外工作的专家提供”。16每位参与者对他们事业的根源都有不同的看法。麦卡锡回忆道，“在场的每个人都非常固执地追求他来之前的想法，据我所知，也没有真正的思想交流。” 17
A passionate advocate of symbolic approaches, the mathematician John McCarthy is often credited with inventing the term “artificial intelligence,” including by himself: “I invented the term artificial intelligence,” he explained, “when we were trying to get money for a summer study” to aim at “the long term goal of achieving human level intelligence.” The “summer study” in question was titled “The Dartmouth Summer Research Project on Artificial Intelligence,” and the funding requested was from the Rockefeller Foundation. At the time a junior professor of mathematics at Dartmouth, McCarthy was aided in his pitch to Rockefeller by his former mentor Claude Shannon. As McCarthy describes the term’s positioning, “Shannon thought that artificial intelligence was too flashy a term and might attract unfavorable notice.” However, McCarthy wanted to avoid overlap with the existing field of “automata studies” (including “nerve nets” and Turing machines) and took a stand to declare a new field. “So I decided not to fly any false flags anymore.”13 The ambition was enormous; the 1955 proposal claimed “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.”14 McCarthy ended up with more brain modelers than axiomatic mathematicians of the sort he wanted at the 1956 meeting, which came to be known as the Dartmouth Workshop.15 The event saw the coming together of diverse, often contradictory efforts to make digital computers perform tasks considered intelligent, yet as historian of artificial intelligence Jonnie Penn argues, the absence of psychological expertise at the workshop meant that the account of intelligence was “informed primarily by a set of specialists working outside the human sciences.”16 Each participant saw the roots of their enterprise differently. McCarthy reminisced, “anybody who was there was pretty stubborn about pursuing the ideas that he had before he came, nor was there, as far as I could see, any real exchange of ideas.”17
就像图灵 1950 年的论文一样，1955 年关于人工智能夏季研讨会的提议现在看来具有令人难以置信的先见之明。麦卡锡、香农和及其合作者提出的研究成为了计算机科学和人工智能领域的主要支柱：
Like Turing’s 1950 paper, the 1955 proposal for a summer workshop in artificial intelligence seems in retrospect incredibly prescient. The seven problems that McCarthy, Shannon, and their collaborators proposed to study became major pillars of computer science and the field of artificial intelligence:
1.“自动计算机”（编程语言）
1. “Automatic Computers” (programming languages)
2.“如何编程让计算机使用语言”（自然语言处理）
2. “How Can a Computer be Programmed to Use a Language” (natural language processing)
3.“神经元网络”（神经网络和深度学习）
3. “Neuron Nets” (neural nets and deep learning)
4.“计算规模理论”（计算复杂度）
4. “Theory of the Size of a Calculation” (computational complexity)
5.“自我完善”（机器学习）
5. “Self-improvement” (machine learning)
6.“抽象”（特征工程）
6. “Abstractions” (feature engineering)
7.“随机性和创造力”（包括随机学习的蒙特卡罗方法）。
7. “Randomness and Creativity” (Monte Carlo methods including stochastic learning).
1955 年，“人工智能”一词是一种愿望，而非对一种方法的承诺。从广义上讲，人工智能既包括通过尝试创造机器智能来发现人类智能的构成，也包括一种哲学上不那么复杂的努力，即让计算机执行人类可能尝试的困难活动。
The term “artificial intelligence,” in 1955, was an aspiration rather than a commitment to one method. AI, in this broad sense, involved both discovering what comprises human intelligence by attempting to create machine intelligence as well as a less philosophically fraught effort simply to get computers to perform difficult activities a human might attempt.
这些愿望中只有少数几个推动了人工智能的发展，而现在，人工智能的代名词就是机器能够从数据中学习。在计算机科学家中，从数据中学习在几代人中都被淡化了。
Only a few of these aspirations fueled the efforts that, in current usage, became synonymous with artificial intelligence: the idea that machines can learn from data. Among computer scientists, learning from data would be deemphasized for generations.
在人工智能发展的前半个世纪，大部分时间都致力于将逻辑与硬编码到机器中的知识相结合。从日常活动中收集的数据几乎不是重点；与逻辑相比，数据的地位微不足道。在过去五年左右的时间里，人工智能和机器学习开始被当作同义词使用；记住它不必这样是一种强有力的思维练习。在人工智能生命的最初几十年里，从数据中学习似乎是一种错误的方法，一种非科学的方法。那些不愿意“直接将知识编程”到计算机中的人会使用这种方法。在数据盛行之前，规则盛行。
Most of the first half century of artificial intelligence focused on combining logic with knowledge hard-coded into machines. Data collected from everyday activities was hardly the focus; it paled in prestige next to logic. In the last five years or so, artificial intelligence and machine learning have begun to be used synonymously; it’s a powerful thought-exercise to remember that it didn’t have to be this way. For the first several decades in the life of artificial intelligence, learning from data seemed to be the wrong approach, a non-scientific approach, used by those who weren’t willing “to just program” the knowledge into the computer. Before data reigned, rules did.
尽管热情高涨，但达特茅斯研讨会上的大多数参与者并没有带来多少具体成果。有一组人与众不同。兰德公司由赫伯特·西蒙领导的团队带来了成果，即自动定理证明器。该算法可以证明基本的算术和逻辑定理。但数学对他们来说只是一个测试案例。正如历史学家亨特·海克所强调的那样，这个团队的起点不是计算或数学，而是研究如何理解大型官僚组织以及解决问题的人的心理。18对于西蒙和纽厄尔来说，人脑和计算机是同一类的问题解决者。
For all their enthusiasm, most participants at the Dartmouth workshop brought few concrete results with them. One group was different. A team from the RAND Corporation, led by Herbert Simon, had brought the goods, in the form of an automated theorem prover. This algorithm could produce proofs of basic arithmetical and logical theorems. But math was just a test case for them. As historian Hunter Heyck has stressed, that group started less from computing or mathematics than from the study of how to understand large bureaucratic organizations and the psychology of the people solving problems within them.18 For Simon and Newell, human brains and computers were problem solvers of the same genus.
我们的立场是，描述解决问题行为的适当方式是用程序来描述：根据生物体能够执行的某些基本信息处理，对生物体在不同环境条件下将做什么进行规范。……数字计算机之所以能出现，只是因为它们可以通过适当的编程来执行与人类解决问题时执行的相同的信息处理序列。因此，正如我们将看到的，这些程序在信息处理层面描述了人类和机器解决问题的过程。19
Our position is that the appropriate way to describe a piece of problem-solving behavior is in terms of a program: a specification of what the organism will do under varying environmental circumstances in terms of certain elementary information processes it is capable of performing. . . . Digital computers come into the picture only because they can, by appropriate programming, be induced to execute the same sequences of information processes that humans execute when they are solving problems. Hence, as we shall see, these programs describe both human and machine problem solving at the level of information processes.19
尽管西蒙和纽厄尔在早期人工智能领域取得了许多重大成功，但他们专注于对人类组织的实际研究。他们他们对人类解决问题的方法很感兴趣，这种方法融合了乔尼·佩恩所说的“二十世纪早期英国符号逻辑与美国超合理化组织的管理逻辑的混合体”。20在采用人工智能这个绰号之前，他们将自己的工作定位为由人类和机器组成的“信息处理系统”的研究，借鉴了当时对人类推理的最佳理解。
Though they provided many of the first major successes in early artificial intelligence, Simon and Newell focused on a practical investigation of the organization of humans. They were interested in human problem-solving that mixed what Jonnie Penn calls a “composite of early twentieth century British symbolic logic and the American administrative logic of a hyper-rationalized organization.”20 Before adopting the moniker of AI, they positioned their work as the study of “information processing systems” comprising humans and machines alike, that drew on the best understanding of human reasoning of the time.
西蒙和他的同事们深入参与了关于人类作为理性动物的本质的辩论。西蒙后来因其关于人类理性局限性的研究而获得了诺贝尔经济学奖。他与战后知识分子一起，致力于反驳这种观点：人类心理应该被理解为对正面和负面刺激的动物反应。和其他人一样，他拒绝了行为主义的观点，即人类几乎自动地受到反射的驱动，学习主要涉及通过这种经验获得的事实的积累。人类的伟大能力，比如说自然语言或做高等数学，绝不会仅仅从经验中产生——它们需要更多的东西。只关注数据就是误解了人类的自发性和智慧。这代知识分子是认知科学发展的核心，他们强调抽象和创造力，而不是对数据（无论是感官数据还是其他数据）的分析。历史学家杰米·科恩-科尔解释说：“学习不仅仅是获取关于世界的事实的过程，更是培养技能或熟练掌握概念工具的过程，然后可以创造性地部署这些工具。” 21对概念的强调是西蒙和纽厄尔的逻辑理论家计划的核心，该计划不只是通过逻辑过程，而是部署类似人类的“启发式”来加速寻找实现目标的方法。乔治·波利亚等学者研究数学家如何解决问题强调了使用启发式方法解决数学问题所涉及的创造力。22因此，数学不是苦差事——它不像做大量的长除法或减少大量数据。它是一种创造性活动——而且，在它的创造者眼中，它是抵御人类极权主义观点的堡垒，无论是左派还是右派。23 （官僚组织中的生活也是如此——在这幅图中，它不一定是苦差事——它可以是一个创造力的地方。只是不要告诉员工这一点。）
Simon and his collaborators were deeply involved in debates about the nature of human beings as reasoning animals. Simon later received the Nobel Prize in Economics for his work on the limitations of human rationality. He was concerned, alongside a bevy of postwar intellectuals, with rebutting the notion that human psychology should be understood as animal-like reaction to positive and negative stimuli. Like others, he rejected a behaviorist vision of the human as driven by reflexes, almost automatically, and that learning primarily concerned the accumulation of facts acquired through such experience. Great human capacities, like speaking a natural language or doing advanced mathematics, never could emerge only from experience—they required far more. To focus only on data was to misunderstand human spontaneity and intelligence. This generation of intellectuals, central to the development of cognitive science, stressed abstraction and creativity over the analysis of data, sensory or otherwise. Historian Jamie Cohen-Cole explains, “Learning was not so much a process of acquiring facts about the world as of developing a skill or acquiring proficiency with a conceptual tool that could then be deployed creatively.”21 This emphasis on the conceptual was central to Simon and Newell’s Logic Theorist program, which didn’t just grind through logical processes, but deployed human-like “heuristics” to accelerate the search for the means to achieve ends. Scholars such as George Pólya investigating how mathematicians solved problems had stressed the creativity involved in using heuristics to solve math problems.22 So mathematics wasn’t drudgery— it wasn’t like doing lots and lots of long division or of reducing large amounts of data. It was creative activity—and, in the eyes of its makers, a bulwark against totalitarian visions of human beings, whether from the left or the right.23 (And so, too, was life in a bureaucratic organization—it need not be drudgery in this picture—it could be a place for creativity. Just don’t tell that to its employees.)
麦卡锡主义与常识
McCarthy and Common Sense
组织逻辑不是约翰·麦卡锡的拿手好戏。逻辑和常识才是。尤其是创建能够结合逻辑和常识来实现日常目标的程序。麦卡锡的逻辑程序遭到了严厉的批评。在 1958 年于伦敦泰丁顿举行的“思维过程机械化”会议上，奥利弗·塞尔弗里奇嘲笑了对演绎逻辑的关注——称“很多胡说八道”是“演绎逻辑是神圣的东西，你可以借用它来做特别神圣的事情，并产生不可侵犯的结果。”为了说明逻辑和日常推理之间的鸿沟，他对女性的工作进行了令人震惊的厌恶。“大多数女性从未推断过，但她们相处得很好，嫁给了幸福的丈夫，养育了快乐的孩子，根本不使用演绎逻辑。”会议上的另一位批评者继续这种令人遗憾的推理，强调女性是如何通过反馈机制而不是逻辑推理来学习的：“如果她以灾难性的方式摔倒了孩子，她就没有机会了，或者她会大叫一声。她通过粗糙的技术很快就学会了如何实现精确控制。这是直接的反馈！如果她试图赢得配偶并尝试了一个没有得到正确结果的举动对此，麦卡锡回应道：“她很快就改变了策略。” 24对女性知识的提及做了大量的论证工作，与后来的女权主义者对人工智能的批判产生了共鸣。25塞尔弗里奇和其他人对普通人的知识和智力感兴趣，无论男女，而麦卡锡作为逻辑传统的继承者，旨在获得公理数学的空灵知识，这里隐含地是像他和他的合作者这样的男性的领域。
Organizational logics were not John McCarthy’s bag. Logic and common sense were. And particularly creating programs that could combine logic and common sense to achieve everyday goals. McCarthy’s logical program met scathing criticism. At the 1958 conference, “Mechanisation of Thought Processes” in Teddington, London, Oliver Selfridge ridiculed the focus on deductive logic—calling “a lot of nonsense” the “notion of deductive logic being something sitting there sacred which you can borrow for particularly sacred uses and producing inviolable results.” To illustrate the gulf between logic and everyday reasoning, he engaged in a shockingly misogynist invocation of the work of women. “Most women have never inferred it, but they get on perfectly well, marrying happy husbands, raising happy children, without ever using deductive logic at all.” Another critic at the meeting continued with this sorry line of reasoning, to underscore how a woman learns through feedback mechanisms, not logical deductions: “If she drops the baby in a disastrous way, she does not get another chance or she gets a great yelp. She learns very quickly by crude techniques of how to achieve precise control. There is direct feedback! If she is trying to win a spouse and tries a move which does not get the right response, she quickly changes her tack.”24 The reference to women’s knowledge did considerable argumentative work, resonating with later feminist critiques of AI.25 Selfridge and others were interested in the knowledge and intelligence of everyday people, men and women alike, whereas McCarthy, an heir to the logical tradition, aimed for the ethereal knowledge of axiomatic mathematics, implicitly here the domain of men like himself and his collaborators.
那么使用大量数据进行的计算又如何呢？它们并没有消失，但它们并不是当时所用的人工智能。1961 年，在一篇具有里程碑意义的文献综述中描述了一些统计方法后，马文·明斯基 (Marvin Minsky) 辩称：“我不认为这种‘增量’或‘统计’学习方案应该在我们的模型中发挥核心作用。”他承认，这些技术“肯定会继续作为我们程序的组成部分出现”，但实际上只是“默认”的。真正的智能存在于其他地方：“一个人越聪明，就越应该能够从经验中学习到一些相当明确的东西；例如，拒绝或接受一个假设，或者改变一个目标。” 26
What of computations using large sets of data? They didn’t disappear, but they weren’t really artificial intelligence as the term was then being used. After describing some statistical methods in a landmark review of the literature in 1961, Marvin Minsky argued, “I am not convinced that such ‘incremental’ or ‘statistical’ learning schemes should play a central role in our models.” He admitted that such techniques “will certainly continue to appear as components of our programs” but really only by “default.” True intelligence lay elsewhere: “The more intelligent one is, the more often he should be able to learn from an experience something rather definite; e.g., to reject or accept a hypothesis, or to change a goal.”26
尽管如此，麦卡锡和其他志同道合的科学家还是将数学和管理模式的推理和行动置于更广泛的潜在人类知识的模仿之上。他们的方法侧重于“对形式符号表示进行操作的程序指令……从 20 世纪 50 年代中期到 80 年代中期，这是人工智能的主导方法（尽管并非唯一方法）。” 27这种人工智能愿景建立在知识层次结构之上；很多东西都可能被视为智能的一部分。这些人工智能的奠基者极大地缩小了人类活动中哪些部分可以被机器模拟，以及他们认为哪些部分是可以模拟的。历史学家乔恩·阿加尔认为，20 世纪中叶的“计算机化”只发生在有现存的物质计算实践的地方，即已经存在的计算、分类和组织业务的方式。28
For all this, McCarthy and other like-minded scientists privileged mathematical and managerial modes of reasoning and acting over the much broader range of potential human knowledge to be emulated. Their approach focused on “programmed instructions operating on formal symbolic representations. . . . From the mid-1950s to the mid-1980s, it was the dominant (though not the only) approach in AI.”27 This vision for AI rested on a hierarchy of knowledge; lots of things could potentially be considered part of intelligence. These foundational figures in AI dramatically narrowed what parts of human activity could plausibly be emulated by machines as well as which parts they thought tractable for doing so. The historian Jon Agar has argued, “computerization” in the middle of the twentieth century “only took place where there were existing material practices of computation” to build upon—already existing ways of counting and classifying and organizing business.28
让机器标准化到足以编程执行逻辑任务绝非易事：20 世纪 50 年代计算的核心是编程语言、编译器和工具的创建，使人们能够编写不依赖于特定机器特性的程序。这项工作最著名的体现是 Grace Hopper 创建的第一个编译器，它使活跃的科学家能够编程并执行逻辑运算和数据处理。29 正如计算机历史学家 Stephanie Dick 以 Simon 和 Newell 为例指出的那样，在实际实现他们的问题求解器时，“程序员必须适应计算机的功能，并在一定程度上放弃模拟人类实践的承诺。” 30在接下来的章节中，我们将看到在实际计算机中实现具有实际限制的挑战如何成为数据科学发展和独特性的核心。
Making machines standardized enough to program to perform logical tasks was no mean feat: central to computing in the 1950s was the creation of programming languages, compilers and tools enabling people to write programs that did not depend on the idiosyncrasies of particular machines. The work, most famously embodied by Grace Hopper’s creation of the first compiler, made plausible machines that active scientists could program and that could perform logical operations and data processing.29 As computer historian Stephanie Dick notes, using the example of Simon and Newell, in actually implementing their problem solver “the programmers had to accommodate the affordances of the computer and, in so doing, abandon to an extent their commitment to simulating human practice.”30 In the chapters that follow, we will see how the challenges of implementing in actual computers with real limitations are central to the development of and distinctiveness of the data sciences.
资助人工智能机构
Funding the AI Establishment
在数据革命之前，资金问题一直困扰着人工智能，当时科技和风险投资公司的雄厚资金打开了人工智能的大门，大大补充了军事和民间政府的资金。从一开始，私人和公共资助者就提出了质疑。当麦卡锡第一次接触洛克菲勒基金会时，基金会的官员们并不热心，直到资历更老的香农加入，才给了一半的资金。在美国，20 世纪 70 年代的大部分资金来自国防部的各个部门。西蒙和纽厄尔在与兰德公司合作并在兰德公司工作，大量资金来自空军和海军研究办公室。国防高级研究计划局 (DARPA) 为麦卡锡以及其他与达特茅斯人物关系密切的研究人员提供了数十年的资助（DARPA 至今仍在资助先进的人工智能，多年来在自动驾驶汽车的开发中发挥了重要作用，就是一个非常明显的例子）。在美国，人工智能彻头彻尾是国家安全国家的产物，是投资潜在军事和商业用途技术的分散战略的一部分——尽管有时距离任何用途都很遥远。随着这笔资金的涌现，大学和国防研究所的研究人员小型社区得以创建，他们主要围绕象征性的人工智能凝聚在一起，对哪些形式的智能值得追求有着狭隘的看法。
Funding long bedeviled AI before the data revolution, when the deep pockets of technology and venture capital firms opened and greatly complemented military and civilian government funds. From the start, private and public funders raised doubts. When McCarthy first approached the Rockefeller Foundation, its officers were unenthusiastic until the far more established Shannon joined, and then only gave half of the monies requested. In the United States, most of the funding during the 1970s came from various facets of the Department of Defense. Simon and Newell, working with and at RAND, drew heavily upon funds from the Air Force and Office of Naval Research. DARPA, the Defense Advanced Research Projects Agency, funded McCarthy for decades, along with a variety of other researchers in close constellation with the Dartmouth figures (and DARPA continues to fund advanced AI to this day, playing an instrumental role for years in the development of self-driving cars, as one highly visible example). In the US, AI was through and through a product of the national security state, part of a diffuse strategy of investment in technologies of potential military and commercial use—though sometimes very distant from any use. With this funding came the creation of a small community of researchers at universities and defense institutes that coalesced largely around a symbolic AI with a narrow purview of what forms of intelligence were worth pursuing.
面对过度承诺和尖锐批评，这些资金时而增多时而减少。1969 年，《曼斯菲尔德修正案》要求军事资金比以前更接近军事潜力，这使政府的慷慨大方受到质疑。1973 年，英国应用数学家詹姆斯·莱特希尔发表了一份对人工智能研究现状的尖锐批评报告。莱特希尔在描述符号人工智能的成功时，几乎毫不掩饰地指出，“在这些抽象游戏情境中解决问题已经产生了许多巧妙而有趣的程序。”这些成功依赖于整合“关于特定问题领域的大量人类知识”。尽管心理学家对此很感兴趣，“但这些程序在实际问题上的表现一直令人失望。” 31虽然这份报告的重要性经常被夸大，但它捕捉到了人们对高度通用的人工智能解决问题形式的热情的下降。
This funding waxed and waned, in the face of great overpromises and bitter criticism. In 1969, the Mansfield Amendment required that military funding have more proximate military potential than before, putting into question much of the government’s largess. In 1973, the British applied mathematician James Lighthill issued a sharply critical report on the state of AI research. Describing the success of symbolic AI, Lighthill noted, with barely disguised condescension, “problem solving in these abstract play situations has produced many ingenious and interesting programs.” These successes rested upon integrating “a really substantial quantity of human knowledge about the particular problem domain.” And for all the interest to psychologists, “the performance of these programs on actual problems has always been disappointing.”31 While the significance of the report is often overstated, it captured the decline in enthusiasm for highly general forms of artificial intelligence problem-solving.
英国广播公司 (BBC) 播出了一场电视辩论，以回应莱特希尔报告，麦卡锡和图灵在布莱切利的合作者米奇作为图灵梦想和新兴领域的捍卫者参与其中。英国的资助减少，而美国资金不足的研究人员的不满情绪加剧，因为他们对人工智能创始人未能兑现承诺感到沮丧。
The BBC broadcast a televised debate in wake of the Lighthill report, in which McCarthy and Michie, Turing’s collaborator at Bletchley, took part as defenders of Turing’s dream and the nascent field. Funding in the United Kingdom waned, and resentment among less well-funded researchers in the States increased, as they voiced frustration with the failed promises of the AI founders.
取而代之的是尝试复制人类智能的更专业形式。
In their stead came attempts to replicate more specialized forms of human intelligence.
专家系统
Expert Systems
到 20 世纪 70 年代中期，人工智能研究已经从试图以一般方式复制人类智能转变为试图复制专家知识。32改变的不仅是代码。关于谁拥有智能以及智能是什么样子的这一概念本身也发生了变化，从一般能力转变为狭隘但深入的专业知识。与其试图复制天才通才，不如复制专门专家。重点仍然是制定规则；然而，不是一般的智能规则，而是伟大专家的具体规则。例如，三位斯坦福大学的主要研究人员得出结论，人类问题解决者的行为“软弱而肤浅，除了人类问题解决者是专家的领域。” 33 1971 年，马文·明斯基和西摩·帕普特认为，“一个非常聪明的人之所以如此，可能是因为他的知识组织知识具有特定的局部特征，而不是因为他‘思考’的整体特征，除了他自我应用知识的效果外，他的‘思考’可能与孩子没有太大区别。” 34
By the mid-1970s, artificial intelligence research had undergone a shift from attempting to replicate human intelligence in a general way to attempting to replicate expert knowledge.32 Not only the code had changed. The very idea of who had intelligence and what that intelligence looked like had shifted, away from generalized capacities to narrow but deep expertise. Rather than attempting to replicate genius generalists, replicate specialized experts. The focus remained on creating rules; however, instead of general rules of intelligence, specific rules of great experts. Three major Stanford researchers, for example, concluded that the behavior of human problem solvers is “weak and shallow, except in the areas in which the human problem-solver is a specialist.”33 In 1971 Marvin Minsky and Seymour Papert argued, “a very intelligent person might be that way because of specific local features of his knowledge-organizing knowledge rather than because of global qualities of his ‘thinking’ which, except for the effects of his self-applied knowledge, might be little different from a child’s.”34
随着对人类的彻底反思，人们对于如何使用机器也进行了彻底的反思：“理解智能的根本问题不在于识别一些强大的技术，而是如何以一种允许有效使用和交互的方式来表示大量知识的问题。” 35那么，挑战就是如何将专业知识转移到计算机中，从而创建“专家系统”。
With this dramatic rethinking about humans came a dramatic rethinking about what to attempt using machines: “The fundamental problem of understanding intelligence is not the identification of a few powerful techniques, but rather the question of how to represent large amounts of knowledge in a fashion that permits their effective use and interaction.”35 The challenge then was how to move specialized expertise into a computer, in the creation of “expert systems.”
显著的成功包括尝试将科学家对有机化学结构的判断形式化，例如由计算机科学家 Edward Feigenbaum、Bruce Buchanan 和生物学家 Joshua Lederberg 合作创建的专家系统 DENDRAL。36这项努力的最高光辉或许来自于 MYCIN，它使识别细菌的过程自动化，以确保医生开出合适的抗生素。37
Notable successes included attempts to formalize the judgment of scientists concerning organic chemical structures, as in the case of the expert system DENDRAL, created by a collaboration between the computer scientists Edward Feigenbaum, Bruce Buchanan, and the biologist Joshua Lederberg.36 The crowning glory of this effort perhaps came with MYCIN, which automated the process of identifying bacteria in order to ensure that physicians prescribe appropriate antibiotics.37
知识获取瓶颈
The Knowledge Acquisition Bottleneck
可惜的是，这些专家系统的创建非常耗费人力。偶然而复杂的医学或工业生产世界与计算机所需的狭隘规则之间存在巨大差距。事实证明，拥有临床知识的专家不会像计算机那样使用有意识的决策规则。弄清专家的规则既困难又非常昂贵，而且规则往往既不简单也不简洁。
Alas, these expert systems proved very labor intensive to create. The gulf between the contingent, complex world of medicine or industrial production and the narrow rules required by computers is vast. Experts with clinical knowledge for navigating it, it turned out, don’t operate with conscious decision rules like those of computers. Figuring out the rules of experts was hard and very expensive, and the rules often proved anything but simple or concise.
因此，到 20 世纪 70 年代初，许多人工智能从业者都在努力克服将人类专业知识转化为“知识库”和正式推理规则的挑战。人工智能研究人员将这一基本困难称为“知识获取瓶颈”。38无论专家在根据感知执行操作或做出判断方面有多优秀，从艺术鉴赏家到物理学家，他们都很难解释自己的专业知识，更不用说将其转化为计算机所需的明确规则了。想想有多少背景信息理解一个简单的菜谱需要理解很多东西。例如，把肉煎成褐色，主要意味着用平底锅加热使其变灰。澳大利亚研究员 J. Ross Quinlan 指出，试图解释其规则的专家“被要求执行他通常不做的任务，例如制定一个主题的综合路线图”。39唐纳德·米奇在 1985 年《专家系统》杂志的一次采访中被描述为“专家系统最杰出的代言人之一”，尽管如此，他“最近发出了警告”，警告说，要成功理解专业知识，我们需要对其本质有不同的看法：“精通不是通过读书获得的——而是通过反复试验和老师提供的例子获得的。这就是人类获得技能的方式。”米奇指出，这需要对人类作为知识者有一个截然不同的概念：
So, by the early 1970s, many AI practitioners struggled to overcome the challenge of converting human expertise into “knowledge bases” and formal inference rules. Artificial intelligence researchers dubbed this fundamental difficulty the “knowledge acquisition bottleneck.”38 However good experts may be at performing actions or making judgments on the basis of sense perceptions, they all, from art connoisseurs to physicists, struggle to explain their expertise, much less to put it into the explicitly stated rules required by computers. Think only of how much background information is required to understand a simple recipe. To brown meat, for example, largely means rendering it gray through heat on a saucepan. The Australian researcher J. Ross Quinlan noted that an expert trying to explain their rules is “called upon to perform tasks that he does not ordinarily do, such as setting down a comprehensive roadmap of a subject.”39 Donald Michie, described in a 1985 interview in the journal Expert Systems as “one of the most prominent spokesmen for expert systems,” was nonetheless “sounding a note of caution . . . recently,” warning that to succeed at understanding expertise we need to have a different vision of its very nature: “Mastery is not acquired by reading books—it’s acquired by trial-and-error and teacher-supplied examples. That is how humans acquire skill.” Michie noted how this required a dramatically different conception of what humans are as knowers:
人们非常不愿意接受这一点。他们的不情愿告诉我们，我们作为有思想的人，更喜欢哲学上的自我形象。它没有告诉我们当老师或大师训练某人时实际上会发生什么。某人必须从例子中重新生成规则，使它们成为他直觉技能的内在组成部分。40
People are very reluctant to accept this. Their reluctance tells us something about the philosophical self-image that we, as thinking beings, prefer. It tells us nothing about what actually happens when a teacher or a master trains somebody. That somebody has to regenerate rules from example to make them an intimate part of his intuitive skill.40
早期的人工智能渴望模拟一般的问题解决；专家系统则试图模拟高度专业的行为；后来的专家系统则建立在对人类知识的不同看法之上：它通常是一种具体实践，很难纳入规则。创建预测专家熟练判断的量化方法被证明是创建以数据为中心的人工智能的核心。然而，如果能做好这一点，就意味着规则的消亡。在试图使算法化时，从数据中生成符号规则关于专家活动，研究人员在 20 世纪 90 年代创造了机器学习的形式，虽然它们在预测方面取得了成功，但未能产生所需的简明规则。正如我们将看到的，简单的规则并不是赢得预测的途径。
Early AI aspired to emulate general problem-solving; expert systems sought to emulate highly expert behavior; later expert systems were built on a different vision of the knowledge of human beings: it’s often an embodied practice, very challenging to put into rules. Creating quantitative ways of predicting the skilled judgment of experts proved central to the creation of data-centric artificial intelligence. Doing that well, however, proved the death of rules. In attempting to make algorithmic the production of symbolic rules from data about expert activity, researchers by the 1990s had created forms of machine learning that, while they succeeded at predicting, failed to produce the desired concise rules. As we will see, simple rules turned out not to be the path to winning at prediction.
尽管专家系统在学术和商业应用中取得了真正的成功，但它们已经变得高度专业化，在面对陌生事物时缺乏人类推理者的弹性特征。41批评者注意到了这一点。有人认为：
Even as expert systems proved to have real success in academic and commercial applications, they had become vastly specialized and lacked the resilience characteristic of human reasoners in the face of the unfamiliar.41 The critics noticed. One argued:
总体而言，我们可以说，专家系统通过大幅缩小人工智能研究的传统目标，以及模糊巧妙的专业编程与适用于各种领域的统一自组织原则之间的区别，增强了其实用性。尽管人们希望它们具有实用性，但这使得它们对未来更深层次的人工智能技术发展的意义完全值得商榷。42
Overall, we can say that expert systems enhance their pragmatic applicability by narrowing the traditional goals of artificial intelligence research substantially, and by blurring the distinction between clever specialized programming and use of unifying principles of self-organization applicable across a wide variety of domains. This makes their significance for future development of deeper artificial intelligence technologies entirely debatable in spite of their hoped-for pragmatic utility.42
令专家系统社区沮丧的是，这些文字的作者杰克·施瓦茨 (Jack Schwartz) 被任命为 DARPA 下属信息系统技术办公室 (ISTO) 主任，而该部门之前（以 IPTO 的名义）为人工智能开发者提供了大量资金。
Much to the dismay of the expert system community, the author of these lines, Jack Schwartz, was appointed director of the Information Systems Technology Office (ISTO) within DARPA, the division which had previously (under the name IPTO) provided copious funding to the developers of AI.
回到布莱切利园，回到数据
Back to Bletchley Park, Back to Data
1959 年，铁幕两边的学者来到英国国家物理实验室，制定机器智能和“自动编程”的议程。这些论文引起了激烈的争论，比如其中一篇评论员打趣说，约翰·麦卡锡的论文“应该发表在《半生不熟的思想杂志》上”。43在这些关于逻辑和视觉机器的梦想中，英国数学家马克斯·纽曼（图灵的前老师，后来的同事）谈到了一个看似平淡无奇的话题，即机械化他所谓的“更复杂的文书流程”，如确定工资和组织图书馆信息。
In 1959 scholars from both sides of the Iron Curtain came to the National Physical Laboratory in the UK to set out the agenda for machine intelligence and “automatic programing.” The papers occasioned bitter disagreement, as when one commentator quipped that John McCarthy’s paper “belongs in the Journal of Half-Baked Ideas.”43 Amid these dreams of logical and seeing machines, the English mathematician Max Newman, a former teacher and later colleague of Turing, spoke on the apparently bland subject of mechanizing what he called “more complicated clerical processes” such as determining wages and organizing library information.
虽然纽曼对此只字未提，但他却在巧妙地借鉴布莱切利园的经验教训。他在那里建立了一个名为“纽曼计划”的机构。44战争期间，纽曼发明了一种通过对大量密文进行大规模统计分析来解密德国密码的技术；不仅如此，纽曼还与许多未来英国人工智能和统计学界的杰出人物一起推动了开创性的巨型专用计算机的发明，以进行这种分析。战后，他在曼彻斯特建立了一个计算机机构，说服了其他布莱切利园的校友，如图灵和统计学家杰克·古德加入他。正如统计学家用生物学和医学例子来洗白布莱切利园和美国国家安全局的数学教训一样，纽曼将密码学的教训推广到大型数据集上。“很明显，”他写道，“大量的数据处理涉及模式识别，以及对模式是否相似的判断。” 45纽曼把在大型数据集中寻找模式作为学习的核心——并指出需要大量的数据存储才能完成这项任务。
Though he could say nothing of it, Newman was subtly channeling the lessons of Bletchley Park, where he had set up an operation known as the Newmanry.44 During the war Newman had devised a technique to decrypt German codes through the large scale statistical analysis of enormous amounts of ciphertext; more than that, Newman had helped spur creation of the pathbreaking Colossus specialized computers to undertake this analysis, alongside many future luminaries of British artificial intelligence and statistics. He built up a computer operation at Manchester after the war, convincing other Bletchley Park alumni such as Turing and the statistician Jack Good to join him. Just as statisticians laundered the mathematical lessons of Bletchley Park and the NSA with biological and medical examples, Newman generalized the lessons of doing cryptography on large data sets. “It is evident,” he wrote, “that a great deal of data-processing involves the recognition of pattern, and judgement as to whether patterns are alike or not.”45 Newman put finding patterns in large data sets at the heart of learning—and noted the need for vast data stores to accomplish the task.
似乎没有充分的理由说明为什么数字计算机不能被编程为高效的学习机器。它要么最初必须被输入大量关于数据之间的相互联系和概率的信息，要么必须被输入大量关于数据之间的相互联系和概率的信息。否则，它就必须通过充当学习机器来获取这些相互联系，至少在与人类一样广泛的背景下，解决与人类一样多的问题。要做到这一点，它必须拥有巨大的存储容量。46
There would seem to be no good reason why a digital computer could not be programmed to be an efficient learner machine. It would either have to be fed initially with a great amount of information about the interconnections and probabilities between the symbols of man, or else it would have to pick these interconnections up by acting as a learner machine in at least as wide a context and for as many problems, as does man. To do this it would have to have tremendous storage capacity.46
美国和英国的人工智能学术界在很大程度上忽视了这些从数据中学习的理念，而这些理念对于图灵和纽曼等布莱切利的居民来说非常宝贵。许多其他在战争、制造业和商业领域工作的人却没有这样做。
The academic AI community in the US and UK largely ignored these ideas of learning from data, so precious to the denizens of Bletchley like Turing and Newman. Many others working at war, manufacturing, and commerce did not.
在情报界，“幕后”的数据研究仍在继续。模式识别正在发展成为一门应用计算统计领域，主要由军方资助，用于识别图像数据中的物体。就在莱特希尔报告对人工智能进行全面阐述的同时， SRI 国际电气工程师理查德·杜达和彼得·哈特合著的《模式分类和场景分析》一书向学生和研究人员介绍了机器学习的基本思想，包括监督和无监督学习框架。47这种范围更窄但最终更强大的人工智能形式非但没有消亡，反而在工业界蓬勃发展，尤其是在与军工关系密切的部门。
Research into data continued “behind the fence” in the intelligence community where pattern recognition was being developed as an applied computational statistical field, funded in large part by the military to identify objects in image data. Around the same time as the Lighthill report filleted AI, the book Pattern Classification and Scene Analysis by the SRI International electrical engineers Richard Duda and Peter Hart introduced students and researchers to what would become fundamental ideas of machine learning, including the supervised and unsupervised learning frame-works.47 Instead of dying, this narrower, but ultimately more powerful, form of AI thrived in industry, particularly in elements with strong military-industrial ties.
从 20 世纪 50 年代开始，数据开始呈爆炸式增长，人们为理解这些数据所付出的努力也随之激增。当时，很少有人将这些努力视为“人工智能”。但正是这种数据驱动的人工智能方法成就了我们现在的一切，无论好坏。
Data exploded from the 1950s, and so did efforts to understand that data. At the time, few thought of these efforts as “AI.” But it was this data-driven approach to AI which has come to make our present, for better and worse, possible.
大数据是2012年的热门IT热词，随着用于控制海量数据的数量、速度和可变性的经济有效的方法的出现，大数据已变得切实可行。
The hot IT buzzword of 2012, big data has become viable as cost-effective approaches have emerged to tame the volume, velocity and variability of massive data.
–Edd Dumbill，“什么是大数据？” 2012
–Edd Dumbill, “What Is Big Data?” 2012
1953 年夏天，在从洛杉矶飞往纽约的航班上，IBM 销售员 R. Blair Smith 发现自己坐在一位衣冠不整的乘客旁边：“他的白衬衫几天前就该换了。他也需要刮胡子。”他邋遢的邻座原来是美国航空公司总裁 CR Smith。Smith 夫妇（高管，不是乐队成员）聊起了航空公司在管理整个网络的预订数据方面遇到的困难以及 IBM 新的数字数据处理工具。IBM 销售员解释说：“我告诉他，有一台计算机可以做的不仅仅是保持单个航班的可用性。”它可以存储乘客的详细数据：“它甚至可以记录乘客的姓名、乘客的行程，以及他的电话号码。CR Smith 先生对此很感兴趣。你知道，他是一位真正的企业家。” 1 IBM 为军方和美国国家安全局工作，致力于开发用于处理IBM 致力于实时数据处理，它正在寻求将这些前沿技术转化为商业应用的机会，最好是利润丰厚的应用。2 IBM希望让主要的商业客户为其新硬件和软件的研发提供资金支持，就像军事和情报部门所做的那样。IBM 试图将为处理潜在敌机的大量实时数据而创建的能力转化为处理潜在客户的大量实时数据的技术。
In the summer of 1953, on a flight from LA to NYC, an IBM salesman, R. Blair Smith, found himself seated next to an unkempt passenger: “his white shirt should have been changed a couple of days ago. He also needed a shave.” His slovenly seatmate proved to be the president of American Airlines, C. R. Smith. The Smiths (the executives, not the band) got to chatting about the airline’s struggle with managing reservations data across its network and IBM’s new digital tools for data processing. “I told him,” the IBM salesman explained, about “a computer that had the possibility of doing more than just keeping availability” on individual flights. It could store granular data on passengers: “It could even keep a record of the passenger’s name, the passenger’s itinerary, and, if you like, his phone number. Mr. C. R. Smith just was intrigued by this. And, you know, he was a true entrepreneur.”1 Working for the military and NSA, IBM was knee-deep in developing new equipment for dealing with real-time data collected by vast networks of sensors; and it was looking for the opportunity to transfer these frontier technologies into commercial applications, ideally very profitable ones.2 The hope was to get major commercial customers to underwrite its R&D on new hardware and software development, just as the military and intelligence services had been doing. IBM sought to transfer capacities created for dealing with large real-time data about potential enemy aircraft into technologies for dealing with large real-time data about potential customers.
“收集数据和填补座位”：这是在这次对话之后开发的美国航空公司 SABRE 系统的描述副标题。3作为十年开发的产品，预订系统是解决分布式数据生产和决策网络实时处理难题的早期商业解决方案。SABRE（半自动商业研究环境）吸取了政府耗资巨大却失败的教训，即建立名为 SAGE（半自动地面环境）的网络化防空系统。SAGE 是在学术界（麻省理工学院）、工业界（IBM）、军方资助的智库（兰德公司）和新独立的空军的共同努力下创建的，涉及自动记录保存、高质量显示和实时联网。
“Collecting data and filling seats”: so read the subtitle to a description of American Airlines’ SABRE System that developed in the wake of this conversation.3 The product of ten years of development, the reservation system was an early commercial solution to the challenging problems involved in real-time processing of distributed networks of data production and decision-making. SABRE (Semi-Automatic Business Research Environment) drew upon the lessons of the government’s spectacularly expensive and failed attempt to build a networked air-defense system called the SAGE (Semi-Automatic Ground Environment). Created at the intersection of academia (MIT), industry (IBM), military-sponsored think tanks (RAND), and the newly independent Air Force, SAGE involved automated record keeping, high-quality displays, and real-time networking.
第二次世界大战后的四十年间，收集公民和消费者数据的数量激增，收集这些数据的不同机构的数量也随之增加。到 20 世纪 40 年代末，负责信号情报的军事部门及其企业军事承包商能够以计算方式处理数据流。十年后，早期的数字计算机 UNIVAC 为美国人口普查提供支持，私营公司为美国海军的密码学家以及新兴的国家安全局提供了“扩大”以前只能通过打孔卡和人工才能完成的工作。1977 年，美国隐私保护研究委员会沉思道：“机构与个人的关系在多样性和集中度上的变化意味着，个人记录现在几乎涵盖了每个人，影响着每个人的生活，从申请个人贷款的企业高管到申请国家信用卡的学校老师，从向当地银行申请支票担保特权的铆钉工到试图为第一套住房筹措家具的年轻夫妇。” 4有关数据积累及其在个人评估中的作用的问题影响远不止狭义的隐私。它提出了关于基于这些数据的决策过程的访问权限和补救手段的基本问题——谁有权根据数据做出决策以及谁可以质疑这些数据的问题。
In the four decades following World War II, the scale of data collected about citizens and consumers skyrocketed, as did the number of different institutions collecting it. By the end of the 1940s, the military branches charged with signals intelligence and their corporate military contractors could process streams of data computationally. A decade later, the early digital computer UNIVAC was powering the US Census, and private companies were providing the US Navy’s cryptographers along with the nascent NSA the ability to “scale up” what had been previously doable only with punch cards and human labor. The US Privacy Protection Study Commission in 1977 mused that the “change in the variety and concentration of institutional relationships with individuals is that record keeping about individuals now covers almost everyone and influences everyone’s life, from the business executive applying for a personal loan to the school teacher applying for a national credit card, from the riveter seeking check-guarantee privileges from the local bank to the young married couple trying to finance furniture for its first home.”4 Questions around the accumulation of data and its role in evaluation of individuals impacted far more than privacy narrowly conceived. It raised fundamental questions about access to the processes of decision-making based on that data and the means for redress—questions about who has the power to make decisions on the basis of data and who can question it.
20 世纪 70 年代左右，数据收集活动呈爆炸式增长，批评者开始提出关于数据收集对隐私和司法公正影响的关键问题。正如我们将看到的，许多法律和政治问题在 20 世纪 80 年代和 90 年代被搁置一旁，许多有关隐私的讨论被缩减为一个虚弱的躯壳，与私人权力问题脱节，关注的焦点是对政府的恐惧，而不是对私营企业的恐惧。从 20 世纪 90 年代中期到 21 世纪，数据使用的扩张速度远远超过了人们对其潜在危害的广泛认识。自 2010 年以来，关于商业数据使用的当前争论表明，隐私和司法公正密切相关这一概念的回归。
Following the explosion of data collection around the 1970s, critics began asking key questions about the effects of data collection on privacy and justice. As we will see, many legal and political questions were set aside in the 1980s and 1990s, and many conversations around privacy were reduced to an emaciated shell, disconnected from questions of private power, and focused on fear of government rather than of private industry. The velocity of the expansion of the use of data far outpaced broad recognition of its potential harms from the mid-1990s through the 2000s. The current debate over the use of commercial data since 2010 sees the return of a conception of privacy and justice as intimately connected.
规模、国家和公司
Scale, States, and Corporations
1950 年 8 月，一篇在美国各地联合发表的揭发丑闻的专栏文章揭露了一群海军军官帮助创建一家公司来执行一项极为机密的项目。很快，“与这笔交易达成一致的海军军官们就成了这家公司高薪的副总裁。”在接受这家公司的“轻松”工作之前，这些军官们曾为美国国家安全局的前身工作；该项目是第一台用于加密的通用电子数字计算机，由新成立的明尼苏达州工程研究协会 (ERA) 公司制造。5不久之后，该公司开始销售该机器的商业版本，但少了一条会泄露其加密用途的关键指令。撇开潜在的不当行为不谈，当时计算领域的大多数发展都来自如此紧密交织的商业和军事工作，这是冷战时期国家驱动资本主义的特点。
In August 1950, a muckraking column syndicated across the United States revealed that a group of Navy officers helped create a company to pursue a most secret project. And that soon, “the same Navy officers who had made the deal turned up as highly salaried vice presidents of the company.” Before taking their “cushy” jobs at the company, the officers worked for the predecessors of NSA; the project was the first general electronic digital computer for cryptography, built by the new Minnesota company Engineering Research Associates (ERA).5 Before long, the company began selling a commercial version of the machine, minus one key instruction that would disclose its cryptographic purposes. Potential impropriety aside, most developments in computing at the time came from such closely intertwined commercial and military work, characteristic of the state-driven capitalism of the Cold War.
在支持能够存储更大数据的新型数字计算机（先是 ERA，然后是 IBM）的开发的同时，美国国家安全局组织了早期的重要会议，鼓励业界开发强大的数据库解决方案。20 世纪中叶，IBM 是商业信息处理领域的主导公司，但它进入数字计算机业务的时间稍晚。虽然我们倾向于认为新型通用电子计算机与打孔卡处理设备截然不同，但它们最初执行的管理任务与旧机器类似，处理管理组织中的类似问题。
While supporting the development of new digital computers capable of larger data storage first with ERA and then with IBM, the National Security Agency organized key early conferences to encourage industry to develop robust database solutions. The dominant company in business information processing in the mid-twentieth century, IBM, got into the digital computer business a bit late. While we would tend to see the new general-purpose electronic computers as radically different from punched card processing equipment, they initially served similar administrative tasks as older machines and dealt with similar concerns in administrative organization.
大数据需要大型基础设施——而这些基础设施必须得到资助、发明和维护。到 20 世纪 50 年代末，美国政府主要通过军方资助了计算机研发费用的一半以上，政府研究人员也密切参与了开发过程。6为模拟原子爆炸和破解密码而建造的成功计算机很快就有了商业版本。即使是像 SAGE 防御系统这样极其昂贵的失败产品，投入大量资金用于开发对后续发展至关重要的技术，包括 CRT 屏幕和网络技术。美国国家安全局可能拥有第一台晶体管计算机7，并且考虑到其对数据存储和处理的需求，它大量承担了存储介质（如自动磁带系统）的开发成本，这些存储介质可以实现对流数据的近乎实时的分析。这些技术一旦商业化，就可以迁移存储在穿孔卡上的数据，并最终允许对这些数据进行新形式的统计分析，以及收集新形式的数据。
Big data needs big infrastructure—and that infrastructure had to be funded, invented, and maintained. Primarily through the military, the US government funded more than half the research and development costs of computers through the end of the 1950s, and government researchers were intimately involved in the development process.6 Successful computers built for simulating atomic explosions and cracking codes quickly became available in commercial versions. Even horrifically expensive flops, like the SAGE defense system, channeled serious money into the development of technologies central to subsequent developments, including CRT screens and networking technologies. The NSA likely had the first transistorized computer7 and, given its demand for data storage and processing, substantially underwrote the development costs of storage media such as automated tape systems that enabled near real-time analysis of streaming data. Once commercialized, these technologies made possible the migration of data stored on punched cards, for example, and eventually allowed for new forms of statistical analysis of that data—as well as for new forms of data to be collected.
因此，数字计算机不仅提高了计算速度，而且对我们的故事来说更重要的是，提高了数据收集、处理和存储的规模。最初，大部分工作涉及将以前可用的信息数字化，但很快，近乎实时地捕获和存储信息的能力将使数据收集和大型行政组织（从航空公司到福利机构）的运营发生根本性转变。这一切都不是不可避免的。虽然回想起来，这些转变中的许多似乎都是显而易见和预先确定的，但它们涉及倡导者和销售人员推动组织参与这些新能力、巨大且始终被低估的技术成本以及机构逻辑的转变。它们涉及对哪些工作和知识重要以及如何改变的选择——或者不改变。军国主义和资本主义不仅仅是导致向计算机数据处理的转变，因为正是这些过程通过使计算机成为其中的核心，改变了军事和资本主义本身的性质。
So digital computers gained both speed in performing computations and, more importantly for our story, scale in collecting, processing, and storing data. Initially much of the work involved digitizing previously available information, but soon the capacity to capture and store information in close to real time would enable radical shifts in data collection and the running of large administrative organizations, from airlines to welfare agencies. None of this was inevitable. While many of these shifts seem obvious and predetermined in retrospect, they involved advocates and salespeople pushing organizations to partake of these new capacities, intense and always understated technological costs, and shifts in institutional logics. They involve choices about what work and knowledge matters and how they should be changed—or not. And militarism and capitalism didn’t simply cause the shift to computer data processing, as these very processes transformed the military and the nature of capitalism itself by making computers central to them.
几十年来，为 NSA 开发的技术背后的推动力一直处于暗中。但有时从军事到商业应用的转移要公开得多。早期的计算机公司吹捧这些潜在的应用，因为它们非常昂贵且维护成本很高机器。1948 年 UNIVAC 的宣传册询问“你的问题是什么？”，并指出了数据处理和计算能力：“是繁琐的记录和商业和工业的艰巨数字工作？还是科学的复杂数学？” UNIVAC 可能适用于“空中交通管制、人口普查、市场研究、保险记录、空气动力学设计、石油勘探、搜索化学文献和经济规划等各种应用。”数据收集正在标准化，宣传的核心是未来降低成本：“自动化操作是处理各种信息时更经济的关键。”存储空间允许“无限期地”保存“大量文件和海量记录”……但在不再需要时可以删除。8该手册非但没有掩盖政府的慷慨捐助和国防开支，反而赞扬了 UNIVAC 与早期功能性计算机的脱钩，尤其是“陆军军械计算机”（ENIAC）以及人口普查和标准局的支持。
The impetus behind the technologies developed for NSA remained in the shadows for decades. But sometimes the transfer from the military to commercial applications was far more open. Early computer companies touted the potential applications for their very expensive and very high maintenance machines. A 1948 brochure for the UNIVAC asked “What’s Your Problem?” and noted both data processing and computational abilities: “Is it the tedious record-keeping and the arduous figure-work of commerce and industry? Or it is the intricate mathematics of science?” The UNIVAC might be applied to “applications as diverse as air traffic control, census tabulations, market research studies, insurance records, aerodynamic design, oil prospecting, searching chemical literature and economic planning.” Data collection was being normalized, and central to the pitch was lower costs in the future: “AUTOMATIC OPERATION is the key to greater economies in the handling of all sorts of information.” The provision for storage allows for “extensive files and voluminous records” that can be kept “indefinitely . . . yet can be erased when no longer needed.8 Far from hiding the government largesse and defense spending that made the machine possible, the brochure celebrates the descent of the UNIVAC from earlier functional computers, especially the “Army Ordnance Computer” (the ENIAC) and the support of the Census and Bureau of Standards.
转换和存储大量数据的残酷现实很快就浮现出来，这提醒我们，数据始终是物质的，其安全性依赖于密集的基础设施，其实现需要大量且往往是隐藏的劳动力。例如，存储磁带很容易折叠，以至于这个问题被隐藏在技术术语“dolf”下，主要通过不太明智的润滑剂使用和卷轴功率调节来解决。9历史学家珍妮特·阿巴特 (Janet Abbate) 强调，围绕新计算技术的宣传低估了使其发挥作用所需的人力。阿巴特认为，关于劳动力节省的计算，例如声称 ENIAC 可以在两小时内完成 25 个人工月的工作，“既不包括编写程序的‘女性工作’，也不包括机器维护的‘男性工作’。” 10
The cold reality of converting and storing large amounts of data soon hit, a sobering reminder that data is always material and rests on dense infrastructures for its security and massive, often hidden labor for its realization. Storage tapes, for example, were so prone to fold that the problem was hidden under the technical term “dolf,” solved largely through a less judicious use of lubricant and tuning of power applied to the reels.9 Historian Janet Abbate has underscored that publicity around new computing technologies downplayed the human labor needed to make them function. Calculations about labor savings, as with a claim that the ENIAC could do twenty-five man-months of work in two hours, Abbate has argued, “included neither the ‘feminine work’ of preparing programs nor the ‘masculine’ work of machine maintenance.”10
UNIVAC 广告。Eckert-Mauchly 计算机公司 (EMCC)，1948 年。计算机历史博物馆档案 X3115.2005。由加利福尼亚州山景城计算机历史博物馆提供。
UNIVAC Advertisement. Eckert-Mauchly Computer Corporation (EMCC), 1948. Computer History Museum Archive X3115.2005. Courtesy of the Computer History Museum, Mountain View, CA.
尽管人们都在谈论“电子大脑”和计算机，但“电子数据处理”仍然是将旧系统转换为大规模数字系统的参考术语。虽然将各种商业、科学和行政工作转换为数字计算机似乎显而易见，但每次转型都涉及有针对性的宣传。11向组织引入新数据处理技术的专家强调需要小心谨慎，以避免员工和管理层的抵制。一本早期的指南包含了这样的智慧：“任何新操作系统的引入都面临着改变人们习惯的最大障碍……人为问题在复杂性和难度上超过了技术问题。” 12 Datamation等杂志上的广告强调了组织面临的挑战。“当有远见的公司转向 EDP [电子数据处理] 簿记时，”施乐公司 1965 年的一则广告指出，“不间断的过渡是一个大问题。即使是最精心策划的转换，也会出现记录丢失、延迟和混乱的情况。” 13控制数据公司在同一期的广告中询问“强大的数据卡是否会像算盘一样走向衰落”，并承诺“新方法可以防止打孔卡和打孔纸带挡在你和计算机之间”。14结果远没有预期的那么具有革命性。《商业周刊》 1958 年的一篇报道指出，工业界“几乎以宗教般的热情采用了极其复杂的电子计算机”，但“似乎常常不知道该如何处理它们”。工业界同样“感到不满，因为早期的结果远远没有达到它们被包装的美好梦想”。然而，这篇文章宣称向这些机器的过渡“不可避免”，因为“计算机仍然是庞大的工业、商业和政府巨头建立新组织系统的关键”。15企业和政府数据库确实蓬勃发展到了 20 世纪 60 年代，需要集中和标准化分散在各个国家的纸条上的数据，这需要艰苦的工作。历史学家保罗·爱德华兹 (Paul Edwards) 在谈到当时气候模型的计算机化时指出：“与所有基础设施项目一样，这些变化不仅涉及科学和技术创新，还涉及制度转型。” 16
For all the talk of “electronic brains” and computers, “electronic data processing” stuck as the term of reference for converting older systems into large-scale digital ones. While converting all sorts of business, scientific, and administrative work to digital computers may seem obvious in retrospect, each transformation involved targeted advocacy.11 Specialists in introducing new data processing technologies to organizations emphasized the care needed to avoid resistance from employees and management alike. One early guidebook included this wisdom: “The introduction of any new system of operation faces its greatest obstacles in changing the habits of people. . . . The human problems exceed the technical problems in complexity and in difficulty.”12 Advertisements in magazines such as Datamation underscored the challenges organizations faced. “When forward-looking companies convert to EDP [electronic data processing] bookkeeping,” a Xerox ad in 1965 noted, “transition without disruption is the big problem. Missing records, delays, and confusion can plague even the most carefully planned changeover.”13 Asking whether “the mighty data card will go the way of the abacus,” an ad from Control Data Corporation in the same issue promised “new ways to keep punched cards and punch tape from coming between you and your computer.”14 And the results were far less revolutionary than expected. A 1958 report in Business Week noted that industry “has adopted the marvelously complex electronic computers with an almost religious fervor,” and yet “often seems unsure of what to do with them.” Industry is likewise “disgruntled, because early results have fallen far short of the rosy dreams in which they came wrapped.” And yet the article nevertheless proclaimed the transition to these machines “inevitable,” as “computers still hold the key to new systems of organization for the sprawling giants of industry, commerce and government.”15 Corporate and government databases were indeed burgeoning by the 1960s, centralizing and standardizing data often collected on slips of papers distributed across countries, requiring arduous work. Writing about the computerization of climate models at this time, historian Paul Edwards notes, “Like all infrastructural projects, these changes involved not only scientific and technological innovation, but also institutional transformation.”16
信用评分提供了一个关键的例子：社会学家玛莎·潘 (Martha Poon) 展示了信用评分是如何开始的，它是一项根据各个公司收集的客户数据量身定制的活动，具有高度具体的信用模型。从交易数据中创建信用数据既费力又费力，通常需要家庭妇女从事家庭手工业，还要经过漫长的打卡过程。17 随着记录的计算机化，出现了新的信用统计模型，正如历史学家乔什·劳尔 (Josh Lauer) 解释的那样，“计算机辅助信用评分促使信用概念和语言发生了根本性转变，甚至比计算机报告的影响更大。除了减少或消除债权人和借款人之间的人际接触外，评分系统还将信用重新定义为抽象统计风险的函数。” 18在对消费者的日益评估中，大规模的消费者数据收集与大规模计算相遇。到 20 世纪 80 年代，计算机建模使信贷行业能够生产出将客户信息商品化的新金融产品。19计算机本身并没有实现这一点：它们的能力是主动参与并转变的，以适应这些机构不断变化的性质。
Credit scoring offers a key case in point: the sociologist Martha Poon has shown how credit scoring began as a bespoke activity tailored to the data individual firms collected on their customers, with highly specific models of creditworthiness. Creating credit data from transactional data was arduous and laborious, often involving women at home in a cottage industry, and long processes of punching cards.17 Following the computerization of records, new statistical models of creditworthiness emerged, as historian Josh Lauer explains, “Computer-assisted credit scoring precipitated a fundamental shift in the concept and language of creditworthiness, even more so than computerized reporting. In addition to reducing or eliminating human contact between creditors and borrowers, scoring systems redefined creditworthiness as a function of abstract statistical risk.”18 Large-scale data collection on consumers met large-scale computation in the increasing evaluation of consumers. By the 1980s, computer modeling allowed the credit industry to produce new financial products that commodified their customers’ information.19 Computers themselves didn’t make this happen: their capacities were proactively engaged and then transformed to accommodate the changing nature of these institutions.
这些新技术的整合并非自动发生；从 20 世纪 60 年代到 90 年代，行业期刊和会议的版面上充斥着重新定义问题并使用存储在新计算机上的数据提供解决方案的宣传。计算机系统——以及提供说服持怀疑态度的管理层和与工人和工会抗争的技巧。1965 年，洛杉矶一家电视台播出了一场辩论“计算机是威胁吗？”，兰德公司的几位杰出人物参与其中，讨论自动信贷和大学录取决定是否“让个人受制于狭隘的机器效率”。Datamation 的编辑认为不是，他引用了“人类的失败和无法量化决策过程的所有要素，这使得目前不可能提供一致、灵活和公平的系统。”设计师应该做得更好。“我们相信，”编辑写道，“对于一些今天留给情感、奇想和机会的问题，我们可以保持理智和有条理。我们认为尝试组织和权衡影响决策的因素是可能的（也是明智的），即使最终决定必须留给极其低效的情感。” 20
Integrating these new technologies did not happen automatically; the pages of the trade journals and conferences from the 1960s to the 1990s teemed with pitches for redefining problems and offering solutions using data stored on new computer systems—as well as providing tips on convincing skeptical management and contending with workers and labor unions. In 1965 an LA television station aired a debate “Are Computers a Menace?” involving several luminaries from RAND discussing whether automated credit and university admissions decisions “leave the individual at the mercy of a narrow machine efficiency.” The editor of Datamation thought not, citing “the human failure and inability to quantify all of the elements of the decision-making processes which makes it—right now—impossible to provide systems which are consistent, flexible and fair.” Designers should try better. “We believe,” the editor wrote, “that it is possible to be sane and organized about some problems which today are left to emotion, whimsy and chance. We think it’s possible (and wise) to try to organize and to weigh factors which affect a decision, even if the final decision has to be left to the marvelously inefficient emotions.”20
信息的价值与隐私的复兴
The Value of Information and the Revival of Privacy
“通过巨型记忆机器的发展，一种完全不同的电子监视和控制已经成为可能，”万斯·帕卡德（Vance Packard）在1964 年的《赤裸社会》（The Naked Society）中写道。21 “迄今为止，个人信息通常被输入超级计算机，用于社会用途或经济或政治吸引力。但这种情况会一直持续下去吗？这个问题可能尤其适用于那些正在建立个人生活累积文件的记忆机器。” 22帕卡德并不是唯一一个这样想的人。1976 年，斯坦顿·惠勒（Stanton Wheeler）解释说：“记录过程本身必须被视为有问题的，我们不仅可以问一个人生活中的事件在什么条件下会成为记录，还可以问它是否记录和分析的权力往往是必要的，但掌握在政府和商业手中也很危险。那些建立数据库的人需要考虑如何确保隐私的破坏对人类有益。有先见之明的帕卡德走在了数据收集和分析发展的前沿。
“A quite different kind of electronic surveillance—and control—has become possible through the development of the giant memory machines,” wrote Vance Packard in The Naked Society in 1964.21 “Thus far, the information about individuals is usually fed into the super computers to serve a socially useful or economically or politically attractive purpose. But will it always be? This might especially be asked concerning those memory machines that are building up cumulative files on individual lives.”22 Packard was hardly alone. In 1976 Stanton Wheeler explained, “The very record-making process itself, then, must be regarded as problematic and we can ask not only for the conditions under which events in a person’s life will become a matter of record, but whether it is legitimate for them to become a matter of record.”23 The power to record and to analyze was often necessary, but also dangerous in government and commercial hands. Those building databases needed to consider how to be certain that the destruction of privacy proved beneficial to mankind. The prophetic Packard was ahead of developments in collecting and in analyzing data.
到 1971 年，个人数据的经济价值变得越来越清晰。“新信息技术似乎催生了一种新的社会病毒——‘数据狂热’，”哈佛大学法学教授阿瑟·米勒写道。“我们必须开始意识到生活在一个将信息视为经济上可取的商品和权力来源的社会意味着什么。” 24
By 1971, the economic value of data on individuals was becoming ever clearer. “The new information technologies seem to have given birth to a new social virus—‘data-mania,’ ” Harvard law professor Arthur Miller wrote. “We must begin to realize what it means to live in a society that treats information as an economically desirable commodity and a source of power.”24
在水门事件和非法国内情报活动曝光之后，共和党参议员巴里·戈德华特和民主党参议员萨姆·埃尔文致力于将控制个人数据确立为每一位美国公民的权利。他们的法案旨在限制联邦和州政府以及私营公司侵犯隐私的行为。该法案提议确保美国人享有以下权利：
In the wake of Watergate and revelations about illegal domestic intelligence activities, Republican senator Barry Goldwater and Democratic senator Sam Ervin aimed to establish control of personal data as a right of every American citizen. Their bill sought to restrict the violations of privacy from federal and state governments as well as private corporations. It proposed to secure the following rights to US persons:
1. 不应存在本身就存在秘密的个人数据系统。
1. There must be no personal data system whose very existence is secret.
2. 必须有一种方法让个人能够查明记录中有关他的信息以及这些信息将被如何使用。
2. There must be a way for an individual to find out that information about him is in a record and how that information is to be used.
3. 如果有关他的信息有误，必须有办法让个人更正。
3. There must be a way for an individual to correct information about him, if it is erroneous.
4. 必须记录对系统中任何个人数据的每次重要访问，包括所有获得访问权限的个人和组织的身份。
4. There must be a record of every significant access to any personal data in the system, including the identity of all persons and organizations to whom access has been given.
5. 必须有办法防止个人在未经其同意的情况下，为某一目的收集的有关他的信息被用于其他目的。25
5. There must be a way for an individual to prevent information about him collected for one purpose from being used for other purposes, without his consent.25
在这个框架下，数据库建设者必须对个人负责。公民应该知道正在收集哪些数据以及谁在出于什么目的使用这些数据。他们应该能够阻止数据的收集和移动。他们雄心勃勃的法案最终缩小到仅涵盖联邦政府机构的数据收集和使用。
In this framework, database builders had to be accountable to individuals. Citizens ought to know what data is being collected and who is using it for what purposes. And they should be able to put the kibosh on the collection and movement of data. Their ambitious bill eventually was narrowed to cover only the collection and use of data by federal government agencies.
20 世纪 60 年代，建立全国性中央联邦数据库的计划引发了巨大的隐私担忧，导致该项目被放弃。26仅仅几年后，一个更加阴险的威胁就显现出来。戈德华特和他的盟友看到了无数小型数据库激增所带来的危险：“[我们] 今天正在私营和政府部门建立独立自动化信息系统的零碎部分，这些系统与目前的综合通信结构模式非常相似。” 27
In the 1960s, plans for a national centralized federal database had stoked tremendous privacy fears that led to the abandonment of the project.26 Only a few years later a more insidious threat was apparent. Goldwater and his allies saw the dangers posed by the explosion of countless smaller databases: “[We] are building today the bits and pieces of separate automated information systems in the private and government sectors that closely follow the pattern to the present integrated communication structure.”27
如今为 Netflix 或 Facebook 等平台提供支持的算法技术在 20 世纪 70 年代还处于起步阶段，但从个人数据进行统计推断的算法技术的效力和危险性已经变得清晰起来。国会议员 Victor Veysey 在 1974 年解释了合法使用和个人控制之间所需的平衡。“有必要开发统计数据来解释不断塑造这个国家文化的社会经济趋势，但需要区分为正当目的而收集的数据和有时使用数据的次要目的。”他继续说，“我们不能严重限制信用评分和人寿保险所提供的合法服务”；“然而，我们必须制定适当的控制措施，以确保有关个人事务的信息不被任意买卖。” 28今天如此熟悉的基本上自由的数据交换并不是自然状态，我们的科学、法律和法规应该考虑到这一点。
The algorithmic techniques that fuel platforms like Netflix or Facebook today were in their infancy in the 1970s, but the potency and dangers of algorithmic techniques of statistical inference from personal data were already becoming clear. Congressman Victor Veysey explained in 1974 the balance between legitimate use and personal control needed. “There is a need to develop statistical data to interpret the socioeconomic trends that continually mold the culture of this Nation, but there is a fine distinction to be drawn between data collected for justifiable purposes and the secondary purposes for which the data is sometimes used.” He continued, “We must not severely restrict the legitimate services performed by” credit scoring and life insurance; “yet, we must develop adequate controls whereby information on an individual[’]s personal affairs cannot be bought and sold indiscriminately.”28 The largely free exchange of data so familiar today was no natural state of affairs, and our science, laws, and regulations ought to take that into account.
此外，20 世纪 70 年代早期的公民自由主义立法者认识到，企业和政府侵犯隐私的行为通常围绕种族、性取向和假定的道德品质问题。侵犯隐私对所有人的影响并不相同：数据的积累导致了后果性且往往带有歧视性的把关行为。
Moreover, civil libertarian legislators in the early 1970s recognized that corporate and government invasions of privacy often circled around questions of race, sexual preference, and putative moral character. Invasions of privacy did not affect all people equally: the accumulation of data enabled consequential and often discriminatory gatekeeping.
在回答参议员山姆·努恩的质询时，学者艾伦·威斯汀明确表示：不要求提供某些个人信息将会付出经济代价。
In response to questioning from Senator Sam Nunn, the scholar Alan Westin was explicit: not asking for certain personal information would come at financial cost.
如果我们每年多花 2 美元保费，这样保险公司就无法根据精算要求将非婚生或同性恋者排除在费率基础之外，我认为这对美国公众来说是一笔可以承受的费用，如果交给他们，他们可能会接受。也就是说，参议员先生，多花 2 美元，就不会有人询问我的性生活，也不会有人向我的邻居核实，也不会有人向我的同事汇报我的性生活。我想很多美国人愿意每年多花 2 美元，也不愿意有人调查和报告他们生活的这一方面。29
if it costs us $2 more in a premium a year so that the companies will not be able to claim on an actuarial basis that they have to exclude persons who live out of wedlock or homosexuals from their rate base, I think that is a bearable cost to the American public and one, if put to them, they would probably accept. That is, Senator, by paying $2 more I won’t have people asking about my sex life and checking with my neighbors and doing reports with my fellow employees as to what my sex life is. I think a lot of Americans would be willing to pay $2 a year more and not have that aspect of their lives investigated and reported on.29
当时和现在一样，信息的自由流通降低了一些财务成本。但对个人生活而言，代价确实非常大。一项重要的非政府研究明确指出：“今天，必须权衡隐私与从中心点收集和提供信息所获得的价值或数据库……多少个人信息值得用信用卡换取便利？” 30
Then as now, the free circulation of information lowered some financial costs. But at a very great cost indeed to personal lives. A key nongovernmental study put it clearly: “Privacy must be weighed today against the value gained from the collection and availability of information at central points or data banks. . . . How much personal information is worth the convenience of a credit card?”30
企业对此给出了明确的答案。随着这项法案的出台，银行、直销商、杂志出版商等行业的抱怨迅速而猛烈。他们几乎异口同声地坚持认为，平衡隐私权和企业所需的“信息自由”意味着优先考虑后者。尤其令人恼火的是，要求人们同意其数据的新用途：
Business had a clear answer. In the wake of the introduction of the expansive bill, the complaints of industry came fast and furious, from banks, direct marketers, magazine publishers. In near unison they insisted that balancing the right to privacy with the “freedom of information” needed for business meant prioritizing the latter. Particularly galling were requirements that people consent to new uses of their data:
我们反对禁止在未经个人事先知情同意的情况下转让个人信息的立法。……现代技术允许信贷发放者通过在线终端设施或电话查询获取信贷信息，从而高效、快速地回应消费者。如果法律阻碍了信息的自由流动，由此产生的低效率必然会转化为行业和消费者的更高成本。31
We object to legislation prohibiting the transfer of information concerning individuals without the prior informed consent to those individuals. . . . Modern technology permits credit grantors to respond to consumers efficiently and rapidly partially by virtue of accessing credit information through on-line terminal facilities or alternatively by telephone inquiries. If the free flow of information is impeded by law, the resulting inefficiencies will necessarily be translated into higher costs to industry and consumer.31
而企业和他们的同路人智囊团则抱怨说，保存有关个人信息使用和传输的详细记录既不切实际又繁琐。像许多其他公司一样，历史悠久的零售商西尔斯 (Sears) 抱怨道：“对西尔斯来说，同样极其昂贵的是，需要保留对系统中任何数据的每次访问和使用的完整和准确记录，包括已获得访问权限的所有个人和组织的身份。” 32
And corporations and their fellow-traveler think tanks complained that keeping detailed records about the use and transfer of personal information was impractical and onerous. Like many other corporations, the storied retailer Sears complained, “Also extremely costly for Sears would be the requirement of maintaining a complete and accurate record of every access and use made of any data in a system, including the identity of all persons and organizations to which access has been given.”32
在 20 世纪 70 年代有关隐私法案的辩论中，商会政客和游说者克服了那些关注公民自由的法案，而雄心勃勃的戈德华特-埃尔文法案则缩小范围，只关注联邦政府对信息的收集和使用。1974 年颁布的《隐私法》试图纠正个人控制信息的利益与联邦政府控制和使用信息的利益之间的平衡。颁布的法案删除了私营部门收集、分发或使用数据的规定，并用一个委员会进一步调查取代了最初设想的严格监管。换句话说，联邦政府没有确认保护个人数据的一般原则，也没有为这些数据的收集、交换和销售提供一般形式的会计处理。相反，继早期的信用信息保护之后，美国人获得了至关重要但较窄的保护，仅限于特定的数据领域，最明显的是学生数据（FERPA，1974 年通过）和整整二十年后的医疗患者数据（HIPAA，1996 年通过）。制定普遍隐私法的改革势头，源于越南战争时期对政府的怀疑、对信用机构的担忧、水门事件以及对美国情报机构的揭露，但被浪费了。33
In the course of the 1970s debates over the privacy bills, chamber of commerce politicians and lobbyists overcame those focused on civil liberties, and the ambitious proposed Goldwater-Ervin bill narrowed to focus exclusively on federal government collection and use of information. The enacted Privacy Act of 1974 sought to redress the balance between the interest of individuals to control information and the interest of the federal government to control and use that information. The enacted bill excised the provisions for the collection, distribution of or use of data in the private sector and substituted a commission to investigate further for the strict regulation initially envisioned. In other words, the federal government affirmed no general principle of the protection of personal data, and it provided no generalized form of accounting for the collection, exchange, and sale of that data. Instead, following the earlier protection of credit information, Americans gained crucial but narrower protections, only within specific domains of data, most notably that of students (FERPA, passed into law 1974) and, a full two decades later, medical patients (HIPAA, passed into law 1996). The momentum of reform to provide a generalized privacy law, born of Vietnam-era suspicion of government, concerns about credit agencies, Watergate, and revelations about US intelligence agencies, was squandered.33
在未能保护非政府数据之后，个人数据的自由使用和滥用似乎成为一种自然状态——不是偶然的，不是会改变的，也不是受我们的政治进程和选择影响的。这种几乎不受限制的数据收集和使用规范为 21 世纪平台利用有关人的细粒度数据以及政府利用商业数据进行大规模监控创造了必要条件。
Following this failure to protect nongovernmental data, the free use and abuse of personal data came to seem a natural state of affairs—not something contingent, not something subject to change, not something subject to our political process and choices. This norm of mostly unrestricted data collection and use created essential conditions for platforms in the 2000s capitalizing on granular data about people and for governments using business data for mass surveillance.
1977 年隐私保护研究委员会承认国会未能解决来自企业部门以及联邦和州官僚机构的危险。《隐私法》的颁布（1974 年）意味着美国没有集中式政府数据库。这些担忧产生了不良的副作用：政府没有建立一个大型数据库来统治我们所有人，而是建立了数百个难以调查的数据库，更不用说监管和监督了，每个数据库都受到不同的监管。规模很重要，因为它极大地改变了原本无害的数据库的隐私影响。随着记录的移动变得越来越顺畅，链接记录的速度越来越快，分析记录组和单个记录的技术也越来越多，网络的发展加剧了这些危险。
The 1977 Privacy Protection Study Commission recognized the failure of Congress to address the danger from the corporate sector as well as the federal and state bureaucracies. The enacting of the Privacy Act (1974) meant no centralized government database in the US. Those concerns had a perverse side effect: rather than one big database to rule us all, the government produced hundreds of databases hard to survey, much less to regulate and police, each subject to different regulation. Scale mattered, for it dramatically changed the privacy implications of otherwise innocuous databases. The growth of networking intensified these dangers, as the movement of records became increasingly frictionless, the speed of linking records increased, and techniques for analyzing groups of records and individual records expanded.
尽管数据收集和分析的速度不断加快，但随后几年几乎没有什么变化。1984 年，隐私权倡导者罗伯特·E·史密斯 (Robert E. Smith) 在美国国会作证时展示了一张图表，说明了教育、零售、医疗和信用评级行业的私人数据库如何与令人眼花缭乱的州和联邦数据库交织在一起。34
Despite the accelerating pace of data collection and analysis, little changed in the years following. In testimony before the US Congress in 1984, the privacy advocate Robert E. Smith presented a diagram illustrating how private databases from educational, retail, medical, and credit rating sectors intertwined with a dizzying array of state and federal databases.34
结合商业和政府数据可以轻松揭露个人生活中令人震惊的部分。史密斯解释了结合商业家庭数据和 IRS 数据的潜力：
Combining commercial and government data could easily reveal startling parts of an individual’s personal life. Smith explained the potency of combining commercial household data and IRS data:
格利克曼先生……国税局现在是否正在租用提供各家庭人口统计资料的计算机化名单，以便他们能够查明我是否去看电影，是否去 Lion D'or 吃晚餐，或是否去拉斯维加斯度过周末，然后确定我是否缴纳了足够的税款？
MR. GLICKMAN. . . . Is the IRS now renting computerized lists that provide demographic profiles of various households so they can find out if I go to the movies, or the Lion D’or for dinner, or Las Vegas for a weekend, and then determine if I am not paying enough in taxes?
史密斯先生：嗯，不完全是那种数据，但它们可以表明你有一辆卡迪拉克和一辆福特。
MR. SMITH. Well, not quite that data, but they could indicate that you had a Cadillac and a Ford.
格利克曼先生：但是他们能查看我的美国运通账户吗？
MR. GLICKMAN. But could they look into, let’s say, my American Express account?
史密斯先生：不，你没有。事实上，你有这样一个账户，你保持的总体平衡，可能在里面，是的。35
MR. SMITH. Not into it; the fact that you had such an account might be reflected, and the general balance that you keep, that might be in there, yes.35
凯瑟琳·麦卡锡 (Kathleen McCarthy)，数据流图，《隐私杂志》，1984 年 4 月。罗伯特·埃利斯·史密斯 (Robert Ellis Smith) 论文，罗伯特·S·考克斯 (Robert S. Cox) 特别收藏和大学档案研究中心，马萨诸塞大学阿默斯特分校图书馆。
Kathleen McCarthy, data flow diagram, Privacy Journal, April 1984. Robert Ellis Smith Papers, Robert S. Cox Special Collections and University Archives Research Center, UMass Amherst Libraries.
上图中的滑轮向国会展示了到 20 世纪 80 年代中期数百个数据库如何整合在一起，创建了一个涉及大多数美国居民的“事实上的”国家数据库。
The pulleys in the diagram above showed Congress how hundreds of databases had come together by the mid-1980s to create a “de facto” national database involving most residents of the United States.
这些组合增强了政府权力。到 20 世纪 80 年代中期，美国国会技术评估办公室报告称，“技术已经改变了数据收集和隐私之间的平衡，使之有利于机构”。将来自多个数据库的数据组合在一起，对隐私产生了重大影响。
These combinations had empowered the government. By the mid-1980s, the Congressional Office of Technology Assessment reported, “Technology has now altered that balance” between collection of data and privacy “in favor of the agencies.” Combining data from multiple databases radically affected privacy.
计算机和电信能力扩大了联邦机构使用和操纵个人信息的机会。例如，为了检测欺诈、浪费和滥用，对存储在不同数据库中的信息进行匹配的次数大幅增加……同样，在个人获得福利、服务或就业之前，计算机越来越多地被用来验证个人信息的准确性和完整性……这些技术能力似乎已经超越了个人保护自身利益的能力。36
Computers and telecommunication capabilities have expanded the opportunities for Federal agencies to use and manipulate personal information. For example, there has been a substantial increase in the matching of information stored in different databases as a way of detecting fraud, waste, and abuse, . . . Likewise, computers are increasingly being used to certify the accuracy and completeness of individual information before an individual receives a benefit, service, or employment. . . . These technological capabilities appear to have outpaced the ability of individuals to protect their interests.36
鉴于计算机能够“匹配”来自不同数据库的记录，如何纠正这种不平衡是一个迫切的问题。1977 年的一个早期例子是“匹配计划”，涉及寻找欺骗福利制度的人。“一个核心政策问题是，考虑到匹配对象的权利以及一般电子搜索可能产生的长期社会影响，使用计算机匹配是否合适以及在什么条件下合适。”毫不奇怪，某些阶层的人更经常受到这种匹配的影响：“计算机匹配本质上是大规模或阶级调查，因为它们是针对一类人而不是特定个人进行的。理论上，没有人能免受这些计算机匹配的影响。”搜查；实际上，福利领取者和联邦雇员最常成为目标。” 37
How to redress the balance given the power of computers to “match” records from different databases was a pressing concern. An early example from 1977, called “Project Match,” involved looking for people cheating the welfare system. “A central policy issue is whether and under what conditions the use of computer matching is appropriate, given the rights of individuals who are the subjects of matching and given the possible long-term societal effects of general electronic searches.” Certain classes of people, to no surprise, find themselves subject to such matching more often: “Computer matches are inherently mass or class investigations, as they are conducted on a category of people rather than on specific individuals. In theory, no one is free from these computer searches; in practice, welfare recipients and Federal employees are most often the targets.”37
数据库技术所能做的远不止在多个数据库中匹配人员。隐私保护研究委员会在 1977 年警告说：“真正的危险是，许多小型、独立的记录保存系统的自动化、集成和互连会逐渐侵蚀个人自由，而这些系统单独来看可能无害，甚至是仁慈的，而且完全合理。” 38规模很重要，它改变了原本无害的数据库的隐私含义。随着记录的移动越来越顺畅，链接记录的速度越来越快，分析记录组和单个记录的技术也越来越多，网络的发展加剧了这些危险。
Database technologies increasingly could do far more than matching people across multiple databases. The Privacy Protection Study Commission warned in 1977: “The real danger is the gradual erosion of individual liberties through the automation, integration, and interconnection of many small, separate recordkeeping systems, each of which alone may seem innocuous, even benevolent, and wholly justifiable.”38 Scale mattered, and changed the privacy implications of otherwise innocuous databases. The growth of networking intensified these dangers, as the movement of records became increasingly frictionless, the speed of linking records increased, and techniques for analyzing groups of records and individual records increased.
隐私和政府利益之间的平衡已经以歧视性的方式倾斜，这种倾斜在今天越来越常见，越来越关注那些最没有能力要求问责、最没有权力反击和要求系统满足我们对公正民主社会中数据分析的全部集体期望的人。39
The balance of privacy and government interest had tilted in discriminatory ways, in ways more and more familiar today, to focus ever more on just those people least able to demand accountability, least empowered to push back and demand that systems fulfill the full range of our collective expectations for data analysis in a just democratic society.39
这些前瞻性的担忧在互联网普及之前就已经出现，当时个人电脑刚刚开始出现在美国及其他国家的家庭和工作场所。随后没有立法。随着数据库的扩大和越来越普遍，日常的数据收集和交换实践巩固了这样一种假设，即除了健康、信用和教育数据的一些主要例外外，没有一般的保护原则来管理非联邦政府和企业对个人数据的使用。缺乏对个人数据的一般保护原则似乎变得越来越自然。缺乏保护个人数据的一般原则并不被视为政治选择，而是隐私保护的本质被错误地理解为数据和数据收集的本质。这一失败的全部意义又花了二十年时间，也就是 2010 年代才被广泛认识到。直到那时，企业和政府贩卖个人数据的双重危险才从少数活动家团体的关注转移到报纸和新闻头版。
The growth of these forward-looking concerns happened before access to the internet became widely available, just as the personal computer was beginning to appear in homes and workspaces across the United States and beyond. No legislation followed. As databases expanded and became ever more ubiquitous, the everyday practice of collection and exchange of data cemented the presumption that no general principle of protection governed non-federal governmental and corporate use of personal data, with some major exceptions for health, credit, and educational data. The lack of any principle of general protection for personal data came to seem ever more natural. Rather than being seen as a political choice, the absence of privacy protection came to be mistakenly understood as the nature of data and data collection. The full significance of this failure took another two decades to become widely apparent, in the 2010s. Only then did the twin dangers of corporate and government trafficking in personal data move from a concern of small groups of activists to the front pages of newspapers and news feeds alike.
1999 年，Sun Microsystems 首席执行官斯科特·麦克尼利坚称：“你反正一点隐私都没有。别在意了。” * 2010 年，Facebook 创始人兼首席执行官马克·扎克伯格声称隐私不再是一种“社会规范”。40这两种说法都不正确。但强大的利益集团却让许多人相信这些观点是正确的。
In 1999, Scott McNealy, the CEO of Sun Microsystems, insisted, “You have zero privacy anyway. Get over it.”* By 2010, Mark Zuckerberg, founder and CEO of Facebook, claimed that privacy was no longer a “social norm.”40 Neither statement is true. But powerful interests worked to make these beliefs seem to many to be true.
1973 年，W. Lee Burge（该公司后来成为今天的消费者信用报告巨头 Equifax）的负责人认为：“通过 [个人和财务] 信息的自由流动——通过掌握准确、相关的事实——美国商人可以而且确实会充满信心地行事，从而保持我们的经济活力和繁荣。” 41这种对数据收集和交换的辩护将创新和经济效率置于其他人类价值之上。在 21 世纪对美国国会对出售美国个人数据的服务的调查的辩护中，主要数据经纪公司 Acxiom 将数据的收集和分析与维护自由本身：“由于许多网络应用程序免费向公众开放，人们每天都会以一种前所未有的方式有机地表达和交流思想。我们最近目睹了各种社交媒体网站如何动员和激励公民……对于这些人群来说，信息确实是通向自由的直接渠道。” 42自由的代价不是警惕，而是被数据挖掘：啤酒的免费再次取代了自由的免费。
In 1973, W. Lee Burge, head of the firm that became today’s consumer credit reporting giant Equifax, argued, “Through the free flow of [personal and financial] information—by having accurate, pertinent facts at their disposal—American businessmen can and do act with the kind of confidence that keeps our economy alive and thriving.”41 Such a defense of the collection and exchange of data prioritizes innovation and economic efficiency over other human values. In a defensive response to a congressional inquiry in the 2000s into services that sell data on US persons, the major data broker Acxiom linked the collection and analysis of data to maintaining freedom itself: “because many web applications are made available free of charge to the public, ideas are organically expressed and exchanged on a daily basis in a manner never before seen. We have recently witnessed how various social media sites have mobilized and energized citizens . . . . For these populations, information truly is a direct conduit to Liberty.”42 The price of freedom is not vigilance but being data mined: free as in beer yet again substitutes for free as in freedom.
从 20 世纪 70 年代至今，自由交换和收集数据的捍卫者们一直在强调，如果我们集体选择更加严格地保护私人数据，那么我们必须预料到一些权衡。他们认为，隐私的代价是巨大的经济成本：服务和产品更加昂贵，创新面临障碍。而另一些人则认为，隐私的代价是国家安全：政府发现和消除邪恶势力的能力下降。这些政治上强有力的论点长期以来一直受到民主党和共和党政府的青睐，它们降低和缩小了我们对个人数据使用及其对隐私、自主权和自由意义的集体期望。
Defenders of the free exchange and collection of data from the 1970s to the present intone about the trade-offs we must expect should we collectively choose to protect private data more robustly. Privacy, they argue, comes at a major financial cost: more expensive services and products and barriers to innovation. And it comes, others argue, at a national security cost: decreased government ability to find and neutralize nefarious forces. These politically powerful arguments, long favored in both Democratic and Republican administrations, lower and narrow our collective expectations about the use of our personal data and its significance for privacy, autonomy, and liberty.
几十年来，业界一直以长期存在的自由市场叙事为依据，强调美国政府在创新方面的缺席，以此来证明我们集体期望值缩小是合理的——尽管正如我们所强调的那样，政府投资创造并培育了计算机行业。在保护个人数据方面，美国在这方面做得很不够。在几年前的一份报告中，一家智库解释道（举个例子）：
Drawing upon long-standing free-market narratives, industry stories for decades have stressed the absence of the US government from innovation to justify the narrowing of our collective expectations—even though, as we have stressed, government investment created and nurtured the computer industry. When it came to protecting personal data, in this narrative, the US has had a light touch. In a report from a few years ago, one think tank explains (to take an example):
在数据经济中，这意味着避免制定限制数据共享和重复使用的全面数据保护规则，而是专注于制定量身定制的特定行业的监管，从而允许大多数行业自由创新。这些政策构成了核心监管环境，使亚马逊、eBay、谷歌和 Facebook 等公司得以蓬勃发展，并为欧洲所采用的预防性、限制创新的规则提供了独特的替代方案。43
In the data economy, this meant avoiding comprehensive data-protection rules that limit data sharing and reuse, and instead focusing on developing tailored regulations for specific sectors, thereby allowing most industries the freedom to innovate. These policies formed the core regulatory environment that allowed companies from Amazon and eBay to Google and Facebook to thrive, and provided a distinct alternative to the precautionary, innovation-limiting rules Europe adopted.43
在这种叙事中，企业自由收集、购买、交易和挖掘数据使美国成为如今的科技强国。在这些故事中，企业的信息流权利压倒了任何透支的隐私权，产生了惊人的效果，并明确表明我们今天应该避免强有力的算法监管和问责形式。这些自由市场故事中缺少的是巨额联邦资金——主要是国防和情报资金——这些资金使微电子行业成为可能并助长了互联网本身。这些故事同样缺少的是我们对少数公司经济效率和利润以外领域的集体合理期望。
In this narrative, corporate freedom to collect, buy, trade, and mine data allowed the US to become the tech powerhouse it is today. In these stories, corporate rights to information flows trumped any overdrawn right to privacy to awesome effect and with the clear moral that we should today avoid robust forms of algorithmic regulation and accountability. Missing from such free-market stories are the vast federal funds—mostly defense and intelligence dollars—that made the microelectronic industry possible and midwifed the internet itself. Equally missing from these tales are our collective legitimate expectations in domains other than economic efficiency and profit of a small number of firms.
1969 年，互联网先驱保罗·巴兰 (Paul Baran) 表示：“我们期望计算机制造商解决隐私问题，可能犯了与期望汽车制造商自行设计出足够的雾霾控制装置一样大的错误。” 44一年前，他在麻省理工学院 (MIT) 的听众面前辩称：“那些处理可以留下烙印和分裂的记录的人必须调整他们的行为，以符合社会的最佳长远利益，即使这种调整与个别机构或公司的最佳利益相冲突。这要求过分吗？” 45
In 1969 internet pioneer Paul Baran remarked: “we may be making as much of a mistake in expecting the computer manufacturers to straighten out the privacy problems as we have made in expecting automobile manufacturers to design adequate smog-control devices of their own accord and without prodding.”44 A year earlier he argued before an audience at MIT: “Those who deal with records that can brand and divide must modify their actions toward the best long-range interests of society, even when such modification conflicts with the best interest of individual agencies or corporations. Is this too much to ask?”45
隐私被削弱
Emaciated Privacy
20 世纪 70 年代，围绕隐私的争论主要集中在自动决策的潜在危害以及对弱势群体可能造成的不成比例的影响上。隐私不仅是原子化个人自由的问题，也是涉及特定人群的公民权利问题，主要收集的关于黑人学生的档案就很好地说明了这一点。20 世纪 70 年代的批评者清楚地看到，规模改变了记录的影响；此后，许多人试图限制我们对这些影响的理解，以使其活动合法化。当时的倡导者强调，隐私概念并非仅仅将隐私视为一项个人权利，而是涵盖了对社会群体之间危害分配不均以及他们寻求正义的能力不同的担忧。到本世纪末，美国关于隐私的广泛讨论大大缩小了，作为个人及其权利自由主义观念更广泛运动的一部分，政治想象力也随之缩小。哈佛哲学家罗伯特·诺齐克曾感叹：“没有任何社会实体会为了自身利益而做出某种牺牲。只有个体，不同的个体，过着各自的生活。” 46经济学家和政策制定者都越来越认同米尔顿·弗里德曼的观点，即社会没有义务，只有个人。47 “在一个建立在私有财产基础上的理想自由市场中，没有个人可以强迫他人，所有合作都是自愿的，所有参与合作的各方都会受益，否则他们就不需要参与。除了个人共同的价值观和责任之外，没有任何价值观，没有任何‘社会’责任。社会是个人及其自愿组成的各个群体的集合。” 48
In the 1970s, arguments around privacy focused squarely on the potential harms of automatic decision making and the likely disproportionate impact upon less empowered groups. Not only a question of liberty of atomized individuals, privacy was a concern of civil rights that pertained to particular classes of people, as dossiers collected primarily on Black students well illustrated. Critics in the 1970s saw clearly that scale altered the effects of records; many people since have tried to limit our understanding of these effects to legitimate their activities. Rather than exclusively considering privacy as an individual right, advocates at the time stressed that the concept encompassed concerns about the uneven distribution of harm among social groups and their different abilities to seek justice. Toward the end of the century broader discussions of privacy in the US narrowed considerably, as did so much of the political imagination as part of the broader movement toward a libertarian conception of the individual and their rights. The Harvard philosopher Robert Nozick exclaimed, “There is no social entity with a good that undergoes some sacrifice for its own good. There are only individual people, different individual people, with their own individual lives.”46 Economists and policymakers alike followed Milton Friedman ever more in his vision that society has no obligation, only individuals.47 “In an ideal free market resting on private property, no individual can coerce any other, all cooperation is voluntary, all parties to such cooperation benefit or they need not participate. There are no values, no ‘social’ responsibilities in any sense other than the shared values and responsibilities of individuals. Society is a collection of individuals and of the various groups they voluntarily form.”48
不仅仅是彻底解除监管国家 在本世纪，这些形式薄弱的社会和经济思维使得人们很难在大数据时代清晰地思考隐私问题。对政府和私人实体的双重强制的关注让位于法律学者乔迪·肖特所说的对政府监管的“偏执风格”。对国家强制危险的深切担忧压倒了对私人权力的担忧。49在这个自由主义的世界里，隐私越来越被狭隘地视为一种个人公民自由，以对抗政府的过度扩张。法律学者普里西拉·里根在 1995 年指出了这种个人主义方法的局限性：“用权利来定义问题一直是许多问题（公民权利、妇女权利、残疾人权利）的强大政治资源，但这些问题涉及获得某种利益或地位的权利，不是以原子个体来定义的，而是以作为群体成员的个人来定义的。” 50虽然激进组织以及奥斯卡·甘迪 (Oscar Gandy) 等杰出批评家极力反对这种批评的狭隘化，但知识趋势和商业利益都削弱了它们在政策甚至激进圈子中应有的突出地位。51这些批评和围绕这些更广泛的隐私概念的技术知识的消失绝非偶然。国会技术评估办公室提供了上述 20 世纪 70 年代的许多见解；纽特·金里奇 (Newt Gingrich) 的众议院于 1995 年关闭了该办公室。就在克林顿总统任期内互联网商业化开始兴起的那一刻，一个政府工作组在刚刚讨论过的 20 世纪 70 年代的工作基础上警告说，互联网将使创建个人资料变得简单而低成本，而无需以前所需的劳动力和旅行。诊断是正确的。然而，媒体学者马修·克雷恩 (Matthew Crain) 认为，当时的治疗方法几乎完全被视为个人选择的问题，赋予个人用户权力。52由此诞生了我们这个无处不在的分析世界，以及个人可以选择退出计算机上的Cookie。53
Not just defanging the regulatory state through the end of the century, these forms of emaciated social and economic thinking made it far less easy to think clearly about privacy in the age of large data. The twin focus on coercion from government and private entities gave way to what the legal scholar Jodi Short called the “paranoid style” of thinking about government regulation. Deep worry about the dangers of state coercion overwhelmed concern over private power.49 In this libertarian world, privacy was increasingly treated narrowly as an individual civil liberty against the overextension of government. The legal scholar Priscilla Regan noted the limits of just such an individualistic approach in 1995: “Defining a problem in terms of rights has been a potent political resource for many issues—civil rights, women’s rights, rights of the disabled—but these issues involve rights to some benefit or status and are defined not in terms of an atomistic individual but an individual as a member of a group.”50 While activist organizations as well as luminous critics like Oscar Gandy pushed hard against this narrowing of critique, intellectual trends and business interests alike muted the prominence in policy and even activist circles that they deserved.51 The loss of these critiques and the technical knowledge around these broader conceptions of privacy was far from accidental. The Congressional Office of Technology Assessment provided many of the insights from the 1970s quoted above; Newt Gingrich’s House of Representatives shuttered the office in 1995. At the very moment when the commercialization of the internet took off during the Clinton presidency, a government working group, building upon the work of the 1970s just discussed, warned that the internet would make the creation of profiles of individuals easy and low-cost, without the labor and travel required previously. The diagnosis was correct. At that moment, the media scholar Matthew Crain argues, however, cures were envisioned almost entirely as a question of individual choice, of empowering individual users.52 And so was begat our world of ubiquitous profiling and of opting out individually of cookies on your machine.53
即使是那些旨在将 20 世纪 60 年代的精神带入互联网的人，也对隐私有着极为个人主义的理解。事实上，政治倡导者庆祝互联网的到来，认为它恰恰破坏了规模差异。埃丝特·戴森解释说，互联网所做的“根本性的事情”是“克服规模经济的优势……这样大人物就无法统治。” 54在这一愿景中，互联网将个人从返祖的社会纽带中解放出来，使弗里德曼的个人主义幻想更加真实，而不是更不真实。历史学家弗雷德·特纳认为，
Even those who aimed to bring the spirit of the 1960s to the internet worked within a dramatically individualistic understanding of privacy. In fact, political advocates celebrated the coming of the internet as quite precisely undermining differences of scale. “The fundamental thing” the internet does, Esther Dyson explained, “is to overcome the advantages of economies of scale . . . so the big guys don’t rule.”54 In this vision, the internet made the individualistic reveries of a Friedman more true not less true by liberating individuals from atavistic social bonds. Historian Fred Turner argues,
尽管凯文·凯利、埃丝特·戴森和约翰·佩里·巴洛等作家描绘出一个无实体的、点对点的乌托邦，但他们却剥夺了许多读者用语言去思考实体如何以复杂的方式塑造人类生活、思考生命所依赖的自然和社会基础设施，以及思考数字技术和网络生产方式可能对生命及其基本基础设施产生的影响。55
Even as they conjured up visions of a disembodied, peer-to-peer utopia, . . . writers such as Kevin Kelly, Esther Dyson, and John Perry Barlow deprived their many readers of a language with which to think about the complex ways in which embodiment shapes all of human life, about the natural and social infrastructures on which that life depends, and about the effects that digital technologies and the network mode of production might have on life and its essential infrastructures.55
在庆祝和捍卫互联网不受政府侵犯的过程中，《连线》等杂志的政治观点强化了隐私是个人权利的狭隘观念。而且，在政治领域中，人们对政府的不信任感非常强烈，认为政府行动迟缓、效率低下。56因此，即使是许多隐私活动人士的言论也让公民严重缺乏应对新聚合带来的风险的能力。互联网在很大程度上使数据和分析成为可能。57尽管许多学者、活动家和技术专家努力拓宽对监控的理解，但在数据收集和分析规模爆炸式增长的当下，政治、社会和法律想象力的这种狭窄化，提供了不足的哲学和法律解释来理解到底发生了什么，也让我们无法想象政治和社会对大数据自动决策的反应是否符合我们的集体愿望。58
In the very process of celebrating and defending the internet against government intrusion, the political vision found in magazines such as Wired reinforced narrow conceptions of privacy as individual rights. And it did so in the broad moment of distrust in governments as slow and inefficient across much of the political spectrum.56 As a result, even many activist accounts of privacy left citizens dramatically underequipped to contend with the risks that the new aggregâtions of data and analysis which the internet largely made possible.57 Despite the work of many scholars, activists, and technologists to broaden the understanding of surveillance, this narrowing of political, social, and legal imagination at just the moment that the scale of data collection and analysis exploded provided inadequate philosophical and legal accounts to understand what had happened and to imagine the political and social response to automatic decision making with large data comport with our collective aspirations.58
2001 年 9 月 11 日恐怖袭击发生前后，美国国家安全局和英国政府通信总部等政府机构已经不再局限于收集和破解纳粹和苏联密码，而是开始收集和分析全球人民（可能包括本国公民）的电话和互联网使用情况。20 世纪 90 年代末，鉴于互联网和移动电话通信的广泛普及，美国和英国的国家安全律师、国防知识分子和执法部门呼吁对窃听相关法律和定义进行改革，但由于公民自由主义者的反对，他们缺乏政治能力来实施这些改革。9/11 事件发生后，美国国会立即通过了 2001 年《爱国者法案》，该法案对国内监控法进行了细微修改，但其全部含义多年来一直不为人知。
Around the time of the attacks of September 11, 2001, government agencies such as the NSA and its British counterpart GCHQ had moved beyond collecting and breaking Nazi and then Soviet codes, as they became able to collect and analyze the telephone and internet use of people worldwide, including potentially their own citizens. National security lawyers, defense intellectuals, and law enforcement in the US and the UK in the late 1990s called for transformation in the laws and definitions around wiretapping, given the vast expansion of communications via the internet and mobile telephony—but lacked the political ability to actualize them over the objections of civil libertarians. In the immediate wake of 9/11, a reactive US Congress passed the PATRIOT Act of 2001, which effected subtle changes in domestic surveillance law, among its many provisions, whose full import remained hidden for years.
由于对隐私的认识不够充分，法官和政策制定者在应对这些新的分析技术时都缺乏想象力。在 9/11 事件之后，本应监管 NSA 的法院表现出了令人震惊的想象力缺乏，他们无法想象规模会如何显著改变数据收集和处理的效果。在 NSA 收集“元数据”的计划曝光之后，高度个人化的隐私方法的局限性很快显现出来。9/11 袭击后，美国国家安全局 (NSA) 一直在收集有关其在 9/11 袭击后所拨打的电话的信息。元数据仅指电话号码，而不指通讯内容。美国国家安全局和其他当局长期以来一直声称元数据并不享有与内容相同的宪法保护，而且他们经常说服立法者和法院同意这一说法。尽管法院有能力要求提供详细说明并有权反击美国国家安全局，但它们缺乏挑战该机构技术主张的能力。美国国家安全局向秘密的外国情报监视法庭 (FISC) 提供了其集体活动的大量正式记录——这是一种闭门透明度，尽管直到最近法庭之外才知道这一点。尽管法庭在法律方面拥有丰富的专业知识，但它缺乏足够的数据聚合技术知识来有效反击这些记录。
Hamstrung by an emaciated view of privacy, judges and policymakers alike suffered from the failure of imagination in contending with these new analytical technologies. The courts that were supposed to regulate the NSA in the wake of 9/11 illustrated a shocking lack of imagination about how scale dramatically changes the effects of the collection and processing of data. The limitations of a highly individualistic approach to privacy appeared quickly in the wake of revelations of the NSA’s program of collecting the “metadata” of phone calls after the 9/11 attacks. Metadata meant only the phone numbers, not the content of the communications. The NSA and other authorities have long claimed that metadata does not share the same constitutional protections as content, and they’ve often convinced legislators and courts to agree. While possessing the ability to demand detailed accounts and empowered to push back against the NSA, the courts lacked the competencies to challenge the technical claims of the agency. The NSA provides vast formal accountings of its collective activities to the secret Foreign Intelligence Surveillance Court (FISC)—a tremendous degree of transparency behind closed doors, albeit unknown outside the court until recently. For all its expertise in the law, the court has lacked sufficient technical knowledge about the aggregation of data to push back effectively against these accounts.
法院的关键判决认定，通信“元数据”受到的宪法保护程度远低于“内容”，这些判决基于一系列关于个人在打电话时缺乏宪法保护的论点。根据普遍的理解，打电话的人会自由地将他们拨打的电话号码告诉电话公司。他们对所拨打的电话号码没有“合理的隐私期望”，尽管他们希望通话内容保持私密。他们拨打的每一个电话都是如此。因此，政府可以获取每个通话的元数据，而不必担心搜查和扣押。缺乏合理的隐私期望从个人通话延伸到对通话的任何汇总或分析。根据这种分析，对没有宪法保护的数据进行操作，只会产生没有宪法保护的事实。
The key court decisions establishing that communications “metadata” have a far lower level of constitutional protection than “content” rest on a series of arguments about the lack of constitutional protection afforded to an individual in the moment of making a telephone call. Under the prevailing understanding, people making telephone calls freely give to the telephone company the phone numbers they are dialing. They have no “legitimate expectation of privacy” as to that dialed phone number, even though they expect the content of the call to remain private. This is true for each call that they make. The government therefore can acquire the metadata for each call without any question of search and seizure. The lack of any reasonable expectation of privacy extends from the individual calls to any aggregation or analysis of them. Operations upon data without constitutional protection, according to this analysis, exclusively yields facts without constitutional protection.
自 2000 年代中期以来，秘密外国情报监视法庭的裁决都是如此处理这一聚合问题显然。第四修正案权利是个人权利：因此“只要没有个人对元数据隐私有合理的期望，那么大量通信将受到……监视的人与是否发生第四修正案搜查或扣押的问题无关。” 59后来的一项裁决进一步阐述了这一推理：“换句话说，如果一个人不具有第四修正案利益，那么将大量处境相似的个人聚集在一起不会导致第四修正案利益凭空而来。 ” 60
Rulings of the secret Foreign Intelligence Surveillance Court since the mid-2000s treat this matter of aggregation plainly. Fourth Amendment rights are personal: so “long as no individual has a reasonable expectation of privacy in meta data [sic], the large number of persons whose communications will be subjected to the . . . surveillance is irrelevant to the issue of whether a Fourth Amendment Search or seizure will occur.”59 A later ruling developed the reasoning further: “Put another way, where one individual does not have a Fourth Amendment interest, grouping together a large number of similarly-situated individuals cannot result in the Fourth Amendment interest springing into being ex nihilo.”60
然而，第四修正案赋予所有个人的利益并非如法院所说的凭空而来；它源于对合法个人隐私利益的挑战，而这种挑战是利用当前的分析工具大规模收集此类数据所导致的。普林斯顿计算机科学家爱德华·费尔滕在一份重要的法庭文件中指出：“复杂的计算工具允许分析大型数据集以识别嵌入的模式和关系，包括个人详细信息、习惯和行为。因此，以前不太可能泄露私人信息的单个数据，现在可能在总体上揭示我们日常生活的敏感细节——我们无意或不期望分享的细节。” 61分析元数据以发现恐怖分子共同模式的前景恰恰取决于这样的假设：这种分析可以发现个人的潜在现象，而不仅仅是总体现象。美国国家安全局自己的历史学家解释了该机构在 1950 年代的能力增长，“除了成功获得邮件内容的底层明文之外，还可以从邮件流量的外部获取有用信息。”这种处理后来被称为元数据的能力“堪称密码学历史上的决定性事件”。62计算统计学的权力，不仅仅是关于集体的统计推断，更重要的是揭露特定个人的私密和个人方面，强调了出于隐私利益的考虑，限制此类分析工具的使用。我们关于在大数据时代是否有能力在知情和理性的情况下同意放弃有关自己的信息的许多旧直觉都是非常错误的。最近关于监控的法院判决，特别是美国诉琼斯案和卡彭特诉美国案，表明司法部门正在慢慢取代一些过时的技术直觉。鉴于我们缺乏关于机器学习平台的权力和危险的认识论和伦理直觉，当前和未来分析工具的权力要求机构具备知识和批判能力，以便在聚合时代重新思考同意问题。63我们的接下来两章将讨论这些强大分析工具的开发。
The Fourth Amendment interest of all the individuals does not, however, arise ex nihilo as the court says; it arises from the challenges to legitimate individual privacy interests that mass collections of such data make possible given current analytical tools. Princeton computer scientist Edward Felten noted in an important court filing, “Sophisticated computing tools permit the analysis of large datasets to identify embedded patterns and relationships, including personal details, habits, and behaviors. As a result, individual pieces of data that previously carried less potential to expose private information may now, in the aggregate, reveal sensitive details about our everyday lives—details that we had no intent or expectation of sharing.”61 The promise of analyzing metadata to discover patterns common to terrorists rests precisely on the assumption that such analysis can uncover latent phenomena about individuals, not just aggregates. The NSA’s own historians explain the growth of the agency’s ability in the 1950s “to derive useful information from the externals of message traffic, in addition to or apart from success in reaching the underlying plaintext of the message contents.” This ability to work with what would later be called metadata “ranks as a defining event in cryptologic history.”62 The power of computational statistics, not just statistical inferences about collectivities, but even more fundamentally to unmask often intimate and personal aspects of specific individuals, underscores that there is a deep privacy interest limiting the use of such analytical tools. Many of our older intuitions about our ability to consent knowingly and rationally to giving up information about ourselves in the era of big data are deeply wrong. Recent court rulings about surveillance, notably United States v. Jones and Carpenter v. United States, show that the judiciary is slowly replacing some of these dated technical intuitions. The power of current and future analytical tools demands institutions with the knowledge and critical power to undertake a rethinking of consent in the age of aggregation, given our lack of epistemic and ethical intuitions about the power—and dangers— of machine learning platforms.63 Our next two chapters consider the development of these powerful analytical tools.
从数据到价值优化
From Data to Optimization as Value
尽管存储数据存在诸多挑战，但事实证明，分析数据以洞察真相要容易得多。在数据收集几乎没有限制的时代，企业和政府数据的增长带来了重大的技术挑战。没有人知道哪些工具可以从这些数据库中产生意义和价值。数据收集的速度越来越快，却没有明确的方法来研究和利用这些数据。
Storing data, for all its challenges, proved far easier than analyzing data for insight. In an era with only minor limits on the collection of data, growing corporate and governmental data presented a major technical challenge. No one knew which tools could produce meaning—and value—from these databases. Data were being collected at an increasing rate without clear means for studying—and profiting—from them.
数据分析技术的资助者已经不耐烦了。几十年来，他们被灌输了许多好东西。例如，美国人口普查局是处理数据技术的早期采用者——从 Hollerith 打孔卡机到上面讨论的 UNIVAC。到了 20 世纪 80 年代，其工作人员对他们资助的那些技术有点恼火。“近三十年来，人口普查局的工作人员一直听到机器识别手写技术即将问世的说法。然而，仔细审查了大多数主张后发现，这一目标还有很长的路要走。” 64由于预算限制越来越紧，以及人们对人工智能和机器翻译等相关领域的夸张主张越来越怀疑，DARPA 和相关机构创建了一种评估项目的新方法，这种方法涉及根据提供给所有竞争对手的数据采用单一指标来评分成功与否。这后来被称为一个通用任务框架。20 世纪 80 年代的发展表明，自动读取手写体技术有理由令人乐观，因此美国人口普查局和美国国家标准与技术研究所 (NIST) 举办了一场比赛，以了解这些长期承诺的技术的现状，鼓励进步，并加强企业和学术界之间的竞争。
Funders of data analysis technologies had become impatient. They had been sold many bills of goods over decades. For example, the US Census Bureau was an early adopter of technologies for contending with data—from Hollerith punched card machines to the UNIVAC discussed above. By the 1980s, its staff had become a bit exasperated with those they were funding. “For almost three decades, staff at the Census Bureau have heard claims that machine recognition of handwriting was just around the technological corner. However, a careful review of most claims showed that the corner was still a long way off.”64 With tighter constraints on budgets, and greater skepticism toward grandiose claims about artificial intelligence and allied fields like machine translation, DARPA and allied agencies created a new approach to evaluating projects, an approach that involved single metrics to score success on data given to all competitors. This became known later as a common task framework. Developments in the 1980s suggested some reason for optimism regarding reading handwriting automatically, so the Census Bureau and the National Institute of Standards and Technology (NIST) set up a competition to see where the long-promised technologies stood, to encourage progress, and to gear up rivalry among firms and academics.
人口普查局不希望竞争对手使用“玩具”数据，因为这些数据与现实世界的数据复杂程度相去甚远。人口普查局的数据是通过数字化手写样本表格获得的，人口普查工作人员和学生将一串数字、字母和单词抄写到明确划分的方框中。
The Census Bureau did not want the competition to work with “toy” data, far removed from the complexity of real-world data. The data were available thanks to a digitized handwriting sample form, where census workers and school children copied out a sequence of numbers, letters, and words into clearly demarcated boxes.
“我们决定，对拥有强大 [光学字符识别] 程序的组织开放测试，这将是实现这些目标的一种经济高效的工具。这将允许比较各种系统、算法、功能和预处理的结果。” 65来自美国和西欧各地的团队加入了竞争——主要来自 Eastman Kodak、Thinking Machines Corporation、IBM 的 Almaden Labs 和戴姆勒-奔驰的 AEG 等公司，还有少数参与者来自从密歇根到瓦伦西亚和博洛尼亚的大学。许多公司都采用了尝试读取地址或支票的商业技术。
“It was decided that a test open to organizations having strong [optical character recognition] programs would be a cost-efficient tool for meeting these goals. This would allow comparison of the results from a wide variety of systems, algorithms, features, and preprocessing.”65 Teams from across the United States and Western Europe joined the fray—mostly from corporations like Eastman Kodak, the Thinking Machines Corporation, IBM’s Almaden Labs, and Daimler-Benz’s AEG, with a smattering of participants from universities from Michigan to Valencia and Bologna. Many of the firms adapted their commercial technologies that attempted to read addresses or checks.
结果如何？“大约一半的系统正确识别了 95% 以上的数字，测试中，人类正确识别了 90% 的大写字母和 80% 以上的小写字母。相比之下，人类正确识别了大约 98.5% 的测试数字。” 66 NIST 公布了结果，如下图所示。
The results? “About half of the systems correctly recognized over 95% of the digits, over 90% of the upper-case letters, and over 80% of the lower-case letters in the test. For comparison, a human correctly recognized about 98.5% of the test digits.”66 NIST made available the results, as seen in the next illustration.
R. Allen Wilkinson、Jon Geist、Stanley Janet、Patrick J. Grother、Christopher JC Burges、Robert Creecy、Bob Hammond 等人，第一届人口普查光学字符识别系统会议。NIST IR 4912，第 19 页。
R. Allen Wilkinson, Jon Geist, Stanley Janet, Patrick J. Grother, Christopher J. C. Burges, Robert Creecy, Bob Hammond, et al., The First Census Optical Character Recognition System Conference. NIST IR 4912, p. 19.
这些系统范围广泛，从神经网络到统计模式识别，从斯皮尔曼的主成分分析到低近邻算法（由统计学家于 1951 年在美国空军的支持下开发）。AT&T 的贝尔实验室提交了四个候选分类系统，包括其商业分类系统的变体基于神经网络的产品。为贝尔实验室的提案做出贡献的研究人员包括许多未来的机器学习杰出人物，包括 Isabelle Guyon 和 Yann LeCun。
The systems ranged widely, from neural networks to statistical pattern recognition, from Spearman’s principal component analysis to the lowly nearest neighbors algorithm (developed by statisticians in 1951, with support of the US Air Force). AT&T’s Bell Labs submitted four candidate classification systems, including a variant of their commercial product based on neural networks. The researchers contributing to the Bell Labs submission included many of the future luminaries of machine learning, including Isabelle Guyon and Yann LeCun.
R. Allen Wilkinson、Jon Geist、Stanley Janet、Patrick J. Grother、Christopher JC Burges、Robert Creecy、Bob Hammond 等人，第一届人口普查光学字符识别系统会议。NIST IR 4912，第 9 页。
R. Allen Wilkinson, Jon Geist, Stanley Janet, Patrick J. Grother, Christopher J. C. Burges, Robert Creecy, Bob Hammond, et al., The First Census Optical Character Recognition System Conference. NIST IR 4912, p. 9.
优化流程以正确分类某些数据与上一章讨论的人工智能宏伟梦想相去甚远。人口普查局和 NIST 坚持的价值观涉及预测的准确性和效率，而不是可理解性或符号逻辑过程的基础；人口普查局和 NIST 同样关心处理大规模现实世界数据的速度。这种价值观的巨大转变是随后机器学习和人工智能爆炸式增长的核心。重视现实世界应用指标的优化推动了 20 世纪 80 年代末至今机器学习、数据挖掘和数据科学的后续发展。准确识别笔迹等问题体现了对具有明确数字成功指标以进行优化的问题的关注。字符识别同样体现了对创建强大的算法系统的坚持，该系统能够处理越来越大规模的现实世界数据，而不是人工清理的数据，通常是实时的。将目标从理解或人工创造“智能”转变为最大化量化性能，也促进了竞争性的社区组织任务。这样的竞赛对于围绕工程性能目标组织社区很有用，但它将焦点从 1956 年达特茅斯研讨会组织者的崇高目标转移了多远。我们的下一章概述了模式识别和机器学习的争议性发展，重点更加明确地关注优化明确成功指标的价值观，而不是广泛但模糊的智能愿望。随后的章节是关于数据科学的，着眼于这些算法如何转变为在工业规模上工作，以处理公司、科学家和政府收集的现实世界数据。这种价值观的缩小优化？这是当今人工智能面临的伦理和政治困境的核心。
Optimizing a process to classify some data correctly was a far cry from the grandiose dreams of artificial intelligence discussed in the previous chapter. The values the Census Bureau and NIST insisted upon involved accuracy of prediction and efficiency, not intelligibility or a grounding in a symbolic logical process; the Census Bureau and NIST were likewise concerned with speed in dealing with its real-world data at scale. Such a dramatic transformation of values is central to the subsequent explosion of machine learning and artificial intelligence. The valuing the optimization of metrics for real world applications drove much subsequent development of machine learning, data mining, and the data sciences from the late 1980s to the present. Problems like the accurate recognition of handwriting exemplify a sharpening of focus on problems characterized by a clear numerical metric of success to optimize. And character recognition likewise exemplifies an insistence of the creation of robust algorithmic systems capable of dealing with real world data, not artificially clean data, at ever-larger scales, often in real time. Changing the goals from understanding or artificially creating “intelligence” to one of maximizing quantitative performance also facilitated a competitive, community-organizing task. Such competitions are useful for organizing a community around an engineering goal of performance, however much it shifts focus away from the loftier goals of, for example, the organizers of the Dartmouth Workshop in 1956. Our next chapter sketches the controversial blossoming of pattern recognition and machine learning focused ever more squarely on values of optimizing for clear metrics of success, rather than broad, but vague, aspirations for intelligence. And the subsequent chapter, on data science, looks at the transformation of these algorithms to work at the industrial scale required to deal with the real-world data collected by corporations, scientists, and governments. And this narrowing of values to optimization? It’s at the heart of the ethical and political dilemmas around AI today.
最初只是用数据进行计算的挑战，后来发展成为一种有利可图的工业化数据狂热，这与麦卡锡和其他人工智能的最初制定者所关注的领域相去甚远——实际上，这对他们来说几乎是诅咒。然而，正如我们将看到的，数据狂热在某种程度上又回来了，为人工智能赋予了第二次或第三次生命——人工智能专注于从数据中学习，而不是手工制定符号规则。
What began as the challenge of computing with data grew into a profitable industrialized data mania, far distant from the concerns of McCarthy and the other original framers of artificial intelligence—indeed nearly anathema to them. Yet, as we will see, the data mania in some ways came back to give a second or third life to artificial intelligence— AI focused squarely on learning from data, not handcrafting symbolic rules.
* Stephen Manes，《私人生活？不是我们的！》，《PC World》，2000 年 6 月。需要澄清的是：当时的首席执行官 Scott McNealy 并没有谈论监控资本主义或机器学习；他谈论的是芯片的设计，相对于竞争对手英特尔奔腾 III 芯片，它可能会暴露计算细节。尽管如此，在接下来的十年里，随着主导技术从硬件公司转向信息平台，这种对消费者保护的漠视和挑衅性在不断重复。Polly Sprenger，《Sun 谈隐私：‘克服它’》，《Wired》，1999 年 1 月 26 日， https://www.wired.com/1999/01/sun-on-privacy-get-over-it/。
* Stephen Manes, “Private Lives? Not Ours!,” PC World, June 2000. To clarify: the CEO at the time, Scott McNealy, was not talking about surveillance capitalism or machine learning; he was talking about the design of a chip, relative to that of the competing Intel Pentium III chip, which could possibly expose the details of computation. Nonetheless, the pithiness and provocativeness of the dismissal of consumer protection was often repeated over the next decade, as the dominant technology changed from hardware companies to information platforms. Polly Sprenger, “Sun on Privacy: ‘Get Over It,’ ” Wired, January 26, 1999, https://www.wired.com/1999/01/sun-on-privacy-get-over-it/.
兰利的P感到失望。到 2011 年，他一生致力于培育的机器学习学术领域在影响力、资金和规模上都出现了爆炸式增长。但成功的代价是巨大的：该领域基本上放弃了“推理、解决问题和语言理解等更复杂的任务”，转而青睐预测等更简单的任务。该领域没有“执行多步推理、启发式问题解决、语言理解或其他复杂认知活动的复杂系统”，而是将自己限制在旨在解决更简单问题的更简单的统计工具上。机器学习已经从模拟人类知识的宏大而有声望的问题转向专注于数值预测和分类。1机器学习似乎更加雄心勃勃，早在四分之一世纪前，也就是 1984 年，他就曾描述过这个领域，将“模式识别”的狭隘目标与人工智能的“符号”方法区分开来：“从历史上看，研究人员对机器学习采取了两种方法。数值方法（如判别分析）在感知领域已被证明非常有用，并且与模式识别范式相关联。相比之下，人工智能研究人员则专注于基于符号学习方法。” 2在此期间，机器学习的价值观发生了变化，其成功的标准也发生了变化。具有讽刺意味的是，正是这种急剧的缩小才使得它今天取得了非凡的成功。20 世纪 70 年代和 80 年代的评论家怀疑人工智能能否成为反乌托邦小说家所想象的那样伟大的东西。21 世纪 20 年代的评论家担心，正如反乌托邦小说家所警告的那样，人工智能将接管人类决策的几乎所有领域。经过最有效的营销人员梦寐以求的重大重塑，“人工智能”一词如今几乎已成为一种更狭义的统计预测技术（称为深度学习）的同义词。本章概述了这个故事。
Pat Langley was disappointed. By 2011, the academic field he’d spent much of his life nurturing, machine learning, had exploded in influence, funding, and size. But success came at a huge cost: the field had largely given up on “more complex tasks like reasoning, problem solving, and language understanding” in favor of simpler tasks like prediction. Instead of “sophisticated systems that carried out multi-step reasoning, heuristic problem solving, language understanding, or other complex cognitive activities,” the field had limited itself to simpler statistical tools designed to solve easier problems. Machine learning had moved from grand, prestigious problems of emulating human knowledge to focus narrowly on numerical prediction and classification.1 Machine learning seemed far more ambitious when he described the same field more than a quarter century earlier, in 1984, separating the narrow goals of “pattern recognition” from the “symbolic” approach of AI: “Historically, researchers have taken two approaches to machine learning. Numerical methods such as discriminant analysis have proven quite useful in perceptual domains, and have become associated with the paradigm known as Pattern Recognition. In contrast Artificial Intelligence researchers have concentrated on symbolic learning methods.”2 In the intervening years, the values of machine learning had changed, as had its criteria for success. Ironically, just this dramatic narrowing has enabled its extraordinary success today. Critics in the 1970s and 1980s doubted artificial intelligence would amount to much of anything, whatever dystopian novelists might envision. Critics in the 2020s worried that artificial intelligence would take over nearly all domains of human decision-making, just as the dystopians warned. Through a remarkable rebranding that the most effective marketers could only dream of, the term “artificial intelligence” today has become nearly synonymous with a narrower slice of statistical techniques for making predictions called deep learning. This chapter sketches that story.
有些领域，比如生物学，以研究对象命名；其他领域，比如微积分，则以方法论命名。然而，人工智能和机器学习是以抱负命名的：这些领域是由目标而不是实现目标的方法定义的。从 20 世纪 60 年代到 21 世纪，机器学习的研究人员从任何必要的地方借鉴方法（尽管遭到许多指定的科学领袖的蔑视）：神经网络、电气工程的“模式识别”方法，甚至数理统计。这些不同的方法论借鉴将继续推动 2010 年代至今的人工智能复兴。
Some fields, like biology, are named after the object of study; others like calculus are named after a methodology. Artificial intelligence and machine learning, however, are named after an aspiration: the fields are defined by the goal, not the method used to get there. From the 1960s to the 2000s, researchers in machine learning took methodologies from wherever necessary (despite the scorn of many anointed scientific leaders): neural nets, the methods of “pattern recognition” from electrical engineering, and even mathematical statistics. These disparate methodological borrowings would go on to forge the AI renaissance of the 2010s to the present.
符号人工智能杀死了神经网络之星
Symbolic AI Kills the Neural Net Star
1980 年，几乎没有人会想到预测模型会接管人工智能。在竞争激烈的融资环境中，符号人工智能的拥护者们嘲笑更多的数据驱动和统计方法。他们尤其贬低了将人类神经网络改造成人工智能的努力。大脑是能够从感知中学习的机器的模型。最著名的例子是感知器，它试图学会区分人工神经元“看到”的物体。
In 1980, few would have expected predictive models to take over AI. Working within a highly competitive funding landscape, the devotees of symbolic AI derided more data-driven and statistical approaches. And they especially denigrated efforts to take the neural networks of a human brain as a model for machines capable of learning from perceptions. The most famous example, the Perceptron, sought to learn to discriminate among objects “seen” by artificial neurons.
感知器是弗兰克·罗森布拉特 (Frank Rosenblatt) 在 20 世纪 50 年代提出的设想，其目的是在没有硬编码规则的情况下识别感官输入。罗森布拉特寻求“一种能够概念化直接来自光、声音、温度等物理环境（光、声音、温度的‘现象世界’）的输入的机器，而不需要人类干预来消化和编码必要的信息。” 3在大量军事资助下，他构建了一个大脑模拟物（一个人工网络）来识别物体，而无需诉诸漫长的逻辑过程。4感知器首先在专用硬件中实现，然后成为能够在通用数字计算机上运行的更标准化的算法。罗森布拉特有宣传天赋，这可能激起了他的批评。 1958 年，《纽约时报》刊登了一篇题为《海军新设备边做边学》的文章，其中的描述令人震惊，描述了“电子计算机的雏形，它有望行走、说话、看、写、自我复制，并意识到自己的存在。” 5
Envisioned by Frank Rosenblatt in the 1950s, the Perceptron involved the effort to recognize sensory input without the hard coding of rules. Rosenblatt sought “a machine which would be capable of conceptualizing inputs impinging directly from the physical environment of light, sound, temperature, etc.—the ‘phenomenal world’ of light, sound, temperature—rather than requiring the intervention of a human agent to digest and code the necessary information.”3 With substantial military funding, he constructed a brain analogue—an artificial network—to recognize objects without resort to long logical processes.4 First realized in specialized hardware, the Perceptron then became a more standardized algorithm able to run on general purpose digital computers. Rosenblatt had a flair for publicity, which may have spurred his critics. In 1958, The New York Times ran a story with the headline “New Navy Device Learned by Doing,” and the claims were stunning, describing “the embryo of an electronic computer that it expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence.”5
该程序是符号人工智能的有力替代品——一种欣赏和模仿人类智能的替代方式——批评者对罗森布拉特的程序进行了激烈的批评。然而，到了 20 世纪 60 年代末，人工神经网络被广泛认为是一条死胡同。简单的神经网络只能使用线性边界对物体进行分类。这是什么意思？为什么这很重要？感知器无法“学习”其输入的一些简单逻辑函数，例如所谓的“排他或”。“排他或”是婚礼中常见的“或”邀请函：你可以选择牛肉、鸡肉或豆腐惊喜，但不能全部选择三样，甚至不能选择两样，除非你从邻居那里偷东西。如果能够进行符号逻辑是智能系统的标志，那么无法处理排他性或就是丧钟。
The program was a potent alternative to symbolic artificial intelligence—an alternate way of appreciating and emulating human intelligence—and critics went after Rosenblatt’s program with great vehemence. By the late 1960s, however, artificial neural networks were widely perceived as a dead end. Simple neural networks can only classify objects using a linear boundary. What does this mean and why does it matter? A Perceptron is not capable of “learning” some easy logical functions of its inputs, such as so-called “exclusive or.” “Exclusive or” is the “or” familiar from wedding invitations: you can choose the beef or the chicken or the tofu surprise, but not all three or even two, unless you steal from your neighbor. If being able to do symbolic logic is the mark of an intelligent system, then being unable to deal with exclusive or is a death knell.
但事实证明，这种限制并没有看上去那么致命。研究人员很快意识到，在第一层之外添加额外的“神经元”层确实可以创建非线性的分类形式。因此，神经网络可以学习“异或”之类的东西。问题是什么？在 20 世纪 60 年代和 70 年代初，没有人知道什么算法可以在大量数据上使用，以有效的方式“训练”多层神经网络，或者以任何程度的信心确保网络会以系统的方式改进。1983 年，经济学家和人工智能先驱赫伯特·西蒙 (Herbert Simon) 自信地宣称，他在 1956 年的达特茅斯研讨会上展示了他的逻辑理论家，
But this limitation proved not quite as a killing strike as it seems. Researchers soon realized that adding additional layers of “neurons” beyond the first layer could indeed create nonlinear forms of classification. Neural nets could thus learn things like “exclusive or.” The rub? In the 1960s and early 1970s, no one knew what algorithm could be used on lots of data to “train” a multilayer neural network in an efficient way or with any degree of confidence that the network would improve in a systematic way. In 1983, the economist and AI pioneer Herbert Simon, who had demonstrated his Logic Theorist at the Dartmouth Workshop of 1956, confidently claimed,
感知器研究和神经网络学习的整个过程……没有取得任何进展……这些系统……从未学习过人们未知的任何东西。因此，它们应该再次加强我们的怀疑，即人工智能的问题只能通过构建学习系统来解决。6
the whole line of Perceptron research and nerve net learning . . . didn’t get anywhere . . . those systems . . . never learned anything that people didn’t already know. So they should again strengthen our skepticism that the problems of AI are to be solved solely by building learning systems.6
神经网络似乎已经死了，死了很久了，除了日本和其他一些地方的少数爱好者。人工智能界的许多人都很高兴没有竞争美国军方的赞助和学术职位。除了金钱竞争之外，神经网络也令人反感，因为它们涉及一种完全不同的智能构成视角。对于像西蒙这样的批评者来说，神经网络的明显缺陷使人们更加怀疑任何试图建立主要关注从数据中学习的学习系统的尝试。
Neural networks seemed dead, way dead, except among a small number of devotees in Japan and a few other places. And many in the AI community were happy not to have competition for US military patronage and academic positions. Besides the competition for dollars, neural nets were distasteful, for they involved a radically different vision of what comprises intelligence. For critics like Simon, the evident failings of neural networks cast doubt more broadly on any attempts to build learning systems focused primarily on learning from data.
然而，远离人工智能的神圣殿堂，经过数据训练的系统并没有消失或失去所有资金。它们被重新定位为当今人工智能的核心。
Away from the hallowed halls of artificial intelligence, however, systems trained on data didn’t disappear or lose all their funding. They are rebranded at the heart of the AI of today.
例如模式识别
Pattern Recognition, for Example
20 世纪 60 年代初，当时刚成为福特汽车公司分支的 Philco 公司工程师与美国陆军签订了合同，共同研究技术手段，帮助军方自动识别 U-2 等侦察机拍摄的照片中的特征。在众多支持技术中，有一项是使用计算统计来帮助对照片中的物体进行分类。正是在美国军方和情报机构资助的这些商业和学术实验室中，计算统计的应用更加侧重于基于数据的预测。Philco 工程师等研究人员在“模式识别”的广泛框架内工作，寻求区分物体的技术，估计已知分布的参数，甚至更具挑战性的是在无法假设其基本形式的情况下开始辨别概率分布的艰巨任务。7他们在政府实验室、企业实验室以及康奈尔大学、南加州大学和斯坦福大学等一流大学工作，通常都得到了军方的大力支持。8没有一家企业实验室能像新泽西州的贝尔实验室那样熠熠生辉。
In the early 1960s, engineers at Philco, newly a division of Ford Motor Company, worked under contract with the US Army on technological means to aid the military in the automated recognition of features in photos taken by spy planes, like the U-2. Among the bevy of technologies supported was the use of computational statistics to aid classification of objects in photos. It was in just such commercial and academic labs, funded by the US military and intelligence agencies, that uses of computational statistics focused more on predictions based on data flourished. Researchers such as the Philco engineers working within the broad rubric of “pattern recognition” sought techniques to discriminate among objects, estimating parameters for known distributions, and, even more challenging, to begin the tough task of discerning probability distributions when their underlying form cannot be assumed.7 They worked at government labs, at corporate labs, and at great universities like Cornell, USC, and Stanford, typically with copious military support.8 No corporate lab shone as brightly as Bell Labs in New Jersey.
当研究人员在 20 世纪 60 年代和 70 年代初期调查该领域时，他们解释说，模式识别涉及的不是一门学科，而是一群志同道合、围绕共同目标的从业者。感知器的神经网络理念可能是这些努力中最著名的。大多数模式识别研究人员最终并不关心——也很少关心——神经网络是否以某种方式复制了人类的认知：网络预测工具，而不是理解大脑的手段。到 20 世纪 60 年代，从业者认为，模式识别之所以成功，很大程度上是因为它放弃了模拟人类感知的努力：“无论我们取得什么成功……都是将感知识别问题有效转化为分类问题的结果。” 9模式识别研究人员很少关心人工智能的符号方面。
When they surveyed the field in the 1960s and early 1970s, researchers explained that pattern recognition involved less an academic discipline than a cluster of like-minded practitioners oriented around common sets of goals. The neural network idea of the Perceptron is perhaps the best known of these efforts. Most researchers in pattern recognition ultimately cared—and care—little whether neural networks in any way replicated human cognition: the networks were tools for prediction, not means for understanding the brain. By the 1960s, practitioners argued, pattern recognition succeeded in large part because it had abandoned the effort to simulate human perception: “Whatever successes we have had . . . have been the result of an effective transformation of a perception-recognition problem into a classification problem.”9 And pattern recognition researchers cared little about the symbolic side of artificial intelligence.
在这些实验室中，一种注重从大量数据积累中获取实际成果的态度盛行起来。在这项工作中，当代数据科学的核心算法的早期形式应运而生，并经过修改，以在当时的计算极限内工作。这意味着，他们不再对符号或模式进行理论化，而是设计出使用真实数据集在有限硬件内实现算法的方法。虽然这些算法出现在学术论文中，但它们主要在实验和商业系统中实现。使用真实数据制作预测系统有时需要丑陋的工程设计。“计算机经济学的实际考虑通常会阻止将上述方法全面应用于现实情况。”这种情况需要“有点不体面和随意的操作……使问题易于有序解决”，包括“预处理、过滤或预过滤、特征或测量提取或降维”。10处理现实世界数据的技术对于实践中的模式识别来说是不可或缺的，而不是辅助的：无论算法多么优雅，如果它无法在有限的磁盘驱动器和计算机上处理来自“现实生活”情况的大规模数据，那么就需要将其搁置或修改。
In these labs, an attitude focused on practical results from large accumulations of data flourished. In the course of this work, early forms of the key algorithms now central to the contemporary data sciences emerged and were modified to work within the computational limits of their times. This meant less theorizing about symbols or schema than devising means for implementing algorithms within limited hardware using real data sets. While these algorithms appeared in academic papers, they were primarily implemented in experimental and commercial systems. Making predictive systems with real-world data required sometimes ugly engineering. “Practical considerations of computer economics often prevent the wholesale application of the methods mentioned above to real-life situations.” Such situations require “somewhat undignified and haphazard manipulation . . . to render the problem amenable to orderly solution,” including “preprocessing, filtering or prefiltering, feature or measurement extraction, or dimensionality reduction.”10 Techniques for handling real-world data were integral, not ancillary, to pattern recognition in practice: no matter how elegant the algorithm, if it could not deal with large-scale data from “real-life” situations on limited disk drives and computers, then it needed to be set aside or modified.
机器学习从模式识别开始学习，甚至更多
Machine Learning Learns, from Pattern Recognition—and More
到二十世纪末，模式识别只是机器学习成功方法的众多来源之一。机器学习本身更多的是一种愿望的规范，而不是一种方法。上述“现实世界”的态度是有代价的。20 世纪 80 年代末和 90 年代的从业者放弃了模拟人类在计算机上推理的方式或使用计算机尝试理解人类认知的人工智能目标。寻求“什么有效”而不是真实或美丽，鼓励人们像喜鹊一样寻找算法和实践来理解数据。机器学习领域缓慢但果断地采用了这些价值观，更多的是实用工程传统而不是纯科学，更符合行业而不是学术界。机器学习研究人员这样做，至少在资金充足的实验室中，他们获得了不均衡但不断增加的计算时间，从而实现了这种方法。11机器学习受到折衷主义的启发，广泛吸收了多个实践和研究领域的算法：模式识别、信号处理、聚类以及以计算为重点的统计学。事实上，统计学家常常抱怨机器学习总是在重复做无用功。回想战时实用统计的传统，自 20 世纪 80 年代末以来，大多数机器学习都涉及最小化某些特定错误或“损失函数”，亚伯拉罕·沃尔德 (Abraham Wald) 将其作为其序贯决策理论的核心，并进入了模式识别领域。12许多机器学习者开始接受贝叶斯统计，尽管情报界人士推崇贝叶斯统计，但长期以来，学术界的数理统计学家一直否认贝叶斯统计。
Pattern recognition was but one of the many sources of successful methods for machine learning by the end of the twentieth century. Machine learning itself was more a specification of an aspiration than of a method. The “real-world” attitudes described above came at a cost. Practitioners over the course of the late 1980s and 1990s abandoned the AI goals of simulating how human beings reason on computers or using computers to attempt to understand human cognition. Seeking out “what works” rather than what is true or beautiful encouraged a magpie-like search for algorithms and practices for making some sense of data. The field of machine learning slowly but decisively adopted these values, more of a practical engineering tradition than of the pure sciences, more aligned with industry than the academy. And machine learning researchers did so with uneven but increasing access to computational time enabling such an approach, at least in well-funded labs.11Animated by eclecticism, machine learning drew widely from algorithms across many fields of practice and inquiry: pattern recognition, signal processing, clustering, as well as from computationally focused statistics. Indeed, statisticians are wont to complain that machine learning keeps reinventing the wheel. Harkening back to the tradition of practical wartime statistics, most machine learning since the late 1980s involves minimizing some specified error or “loss function,” which Abraham Wald put at the heart of his sequential decision theory and made its way into pattern recognition.12 Many machine learners came to embrace Bayesian statistics, long disavowed by mathematical statisticians within the academy, though celebrated within the corridors of the intelligence community.
具有深刻讽刺意味的是，机器学习，一个不受尊重的作为人工智能的近亲，机器学习将在新千年取得最大的成功，甚至是人工智能的救星，以至于在 2013 年之后，机器学习在很大程度上取代了传统人工智能更为宏伟的目标，并且这两个术语开始互换使用。
In a moment of profound irony, machine learning, a little-respected relative of artificial intelligence, would come in the new millennium to become the greatest success, even savior of AI, to such an extent that after 2013 machine learning came largely to displace the far more ambitious goals of traditional AI, and the terms came to be used interchangeably.
从人工智能到机器学习
From Artificial Intelligence to Machine Learning
约翰·麦卡锡 (John McCarthy) 于 1955 年提出的资助提案赞扬了“这样一种猜想：学习的每一个方面或智能的任何其他特征在原则上都可以被如此精确地描述，以至于机器可以对其进行模拟。” 13事实证明，实现人工智能的问题远比瞄准雷达或识别坦克等任务更棘手。詹姆斯·莱特希尔爵士 (Sir James Lighthill) 在 1973 年发表了一份极具批判性的报告，解释道：“人们在 1950 年左右甚至 1960 年左右进入了 [人工智能] 领域，他们寄予厚望，但 1972 年的希望还远远没有实现。迄今为止，该领域的任何发现都没有产生当时承诺的重大影响。” 14人工智能研究既包括尝试创造智能行为，也包括更好地理解人类智能。到 1987 年，一位评论员指出：“没有人再谈论复制人类智能的全部范围。相反，我们看到的是退回到专门的子问题。” 15许多方面已经放弃了人工智能的崇高目标。
John McCarthy’s 1955 funding proposal celebrated “the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.”13 The problems in realizing artificial intelligence proved far more intractable than tasks such as targeting radar or identifying tanks. “Workers entered the field [of artificial intelligence] around 1950, and even around 1960,” the highly critical report by Sir James Lighthill from 1973 explained, “with high hopes that are very far from having been realized in 1972. In no part of the field have the discoveries made so far produced the major impact that was then promised.”14 AI research had included both attempting to create intelligent behavior and better understanding human intelligence. By 1987, a commentator noted, “No one talks about replicating the full gamut of human intelligence any more. Instead we see a retreat into specialized subproblems.”15 The lofty goals of artificial intelligence had been abandoned in many quarters.
承诺与现实之间的差距导致了不止一次“人工智能寒冬”——这是一个经常用来形容政府大笔资金枯竭的季节性比喻。20 世纪 70 年代和 80 年代初出现了第二个兴衰周期，这次是在专家系统中。此类系统的构建者旨在从人类专家那里收集信息，将这些信息组织成系统程序，然后在计算机上实施这些程序，以执行以下任务：医学诊断。尽管在少数领域取得了有限的成功，并悄无声息地融入了许多日常系统，但这些系统却很脆弱，随着 20 世纪 80 年代的结束，它们的市场也陷入了低迷。16事实证明，像赫伯特·西蒙这样的杰出人物的信心是错误的。
The observed gulf between promises and reality led to more than one “AI winter”—a seasonal metaphor constantly used for the drying up of lavish government funding. The 1970s and early 1980s saw a second boom-and-bust cycle, this time in expert systems. The builders of such systems aimed to collect information from human experts, organize that information into systematic procedures, and then implement those procedures on computers, to undertake tasks such as medical diagnosis. Despite limited success in a few domains and the silent integration in many everyday systems, the systems proved brittle and the market for them tanked as the 1980s came to a close.16 The confidence of luminaries like Herbert Simon proved to be misplaced.
在那些试图重新定位人工智能的人看来，基于规则的人工智能的整个项目都建立在对人类知识的误解之上：它不是用简单的规则就能轻易表达出来的。它不是书本上的知识——它更像是一种实践技能。然而，摆脱这种误解使得研究人员能够研究专家的活动，而不必试图了解他们自己是如何做出判断的。1993 年的一篇文章指出，“机器学习算法不是向专家询问领域知识，而是观察专家的任务，并归纳出模仿专家决策的规则。” 17与那些试图模仿人类决策方式的更雄心勃勃的人工智能不同，这种算法的制造者认为它们的行为与人脑完全不同。这种机器学习不是专注于逻辑，也不是采访专家，而是更集中地关注数据、关于人类的数据和人类部分分类的数据。它使用符号人工智能社区在很大程度上回避的模式识别、统计和神经网络工具来实现这一点。大部分工作都是在工业实验室里进行的，这些实验室具有工程思维、雄厚的资金，并且可以使用大量昂贵的计算机时间，比如贝尔实验室和 IBM。
In the eyes of those seeking to reorient artificial intelligence, the entire project of rules-based artificial intelligence rested on a misconception about human knowledge: it wasn’t articulable easily in simple rules. It wasn’t bookish knowledge—it was more like a practiced skill. Escaping this misconception, however, enabled a research effort to study the activities of experts without attempting to understand how they themselves make judgments. An article in 1993 noted, “Rather than asking an expert for domain knowledge, a machine learning algorithm observes expert tasks and induces rules emulating expert decisions.”17 Unlike more ambitious forms of artificial intelligence seeking to emulate how humans make decisions, the makers of such algorithms viewed them as acting in no way like human brains. Rather than a focus on logic, or interviewing experts, machine learning of this sort focused even more centrally on data, data about humans and data partially classified by humans. And it did so using the tools of pattern recognition, statistics, and neural networks that the symbolic artificial intelligence community had in large part shunned. Much of this work happened at industrial labs with an engineering mindset, deep pockets, and access to large amounts of expensive computer time, places like Bell Labs and IBM.
美国和英国以外
Outside the US and UK
在英语世界之外，数据驱动的计算统计学也发展起来，与数理统计和符号人工智能相对立。在法国，让-保罗·本茨克里创建了一个一个强大的“数据分析”学派专注于利用计算机进行更强大的探索性和描述性统计。他写道：“计算机带来的‘数据分析’进步将无法继续，否则整个统计学都会被颠覆。” 18在日本，Hayashi Chikiō 开发了一套实践，他称之为“数据科学”，作为数理统计的替代品，他将其描述为“一无是处，难以理解” 。19
Data-driven computational statistics also developed in opposition to mathematical statistics and symbolic AI outside the Anglophone world. In France, Jean-Paul Benzécri created a powerful school of “analyse des données” [analysis of data] focused on more powerful exploratory and descriptive statistics using computers. “The progress of the ‘analyse des données’ due to computers,” he wrote, “will not continue without upsetting all of statistics.”18 In Japan, Hayashi Chikiō developed a set of practices he named Deta no Kagaku, the “Science of Data,” as an alternative to mathematical statistics, which he described as “good-for-nothing and not understandable.”19
苏联的发展可能对数据及其分析的近期历史影响最为深远。2006 年，机器学习专家 Vladimir Vapnik 反思了几十年前苏联计算机学习的变革。Vapnik 和志同道合的同事拒绝了主流统计方法，创建了“预测（判别）归纳模型”。在这种方法中，“预测模型不一定将事件的预测与对事件规律的理解联系起来；它们只是在寻找一种能够最好地解释数据的函数。” 20 Vapnik 在 20 世纪 60 年代和 70 年代担任苏联科学院控制科学研究所的成员时，接触到了这种工具主义方法和高维数据集。21 虽然针对他和他的拒绝者顾问的反犹太主义可能阻碍了他的职业生涯，但在该研究所的工作让 Vapnik 能够参与到一种高度计算为中心的学习方法的蓬勃发展中，该方法应用于大型数据集。瓦普尼克移居美国，在贝尔实验室工作。在美国和苏联，模式识别和控制理论研究人员都认为自己与符号人工智能和经典学术统计学相距甚远。
Developments in the Soviet Union were probably of the most consequence to the recent history of data and its analysis. In 2006, the machine learning specialist Vladimir Vapnik reflected on the transformations in computer learning in the USSR decades prior. Rejecting dominant statistical approaches, Vapnik and like-minded colleagues created “predictive (discriminative) models of induction.” In such an approach, “predictive models do not necessarily connect prediction of an event with [an] understanding of the law that governs the event; they are just looking for a function that explains the data best.”20 Vapnik came to this instrumentalist approach, and the high-dimensional data sets, as a member of the Institute of Control Sciences of the Academy of Sciences of the USSR in the 1960s and 1970s.21 While anti-Semitism directed toward him and his refusenik advisor likely stymied his career, being in the institute allowed Vapnik to participate in the flourishing of a highly computationally focused learning approach applied to large sets. Vapnik moved to the United States and worked at Bell Labs. In the US and USSR alike, pattern recognition and control theory researchers viewed themselves as distant from symbolic artificial intelligence and from classical academic statistics.
所有这些趋势在 20 世纪 90 年代得到了贝尔实验室的资金和精神支持。当时，贝尔实验室聘请了一批令人印象深刻的国际研究人员，他们开创了机器学习的新方法和新分支，包括未来的杰出人物 Yann LeCun、Yoshua Bengio、Rich Sutton、Rob Schapire 等。与 Vapnik 最相关的技术称为支持向量机 (SVM)，它在一次非凡的合作中结出硕果，当时他与法国研究员 Isabelle Guyon 联手。与计算数据科学领域的其他主要发展案例一样，Vapnik 的工作要求是在支持它的资金机制内处理高维数据，并且没有生产符号人工智能的负担。22尽管它很重要，但贝尔并不是唯一一个拥有如此巨大可能性的地方。在 IBM ，斯坦福大学教授李晓昌表明，统计知识、工程思维、大量语音数据和计算能力的类似融合使得语音识别发生了巨大的转变。23这些工业站点预示并实现了更大的发展。
All these tendencies found financial and moral support at Bell Labs in the 1990s. At that time, Bell Labs had hired an impressive array of international researchers who pioneered new methods and branches of machine learning, including future luminaries Yann LeCun, Yoshua Bengio, Rich Sutton, Rob Schapire, and others. The technique most associated with Vapnik, called support-vector machines (SVMs), came to fruition in a remarkable collaboration there, where he joined forces with the French researcher Isabelle Guyon. Like other major examples of development in the computational data sciences, Vapnik worked under the imperative of contending with high-dimensional data within a funding regime supporting it, and without the burden of producing symbolic artificial intelligence.22 For all its importance, Bell was not the only site for such dramatic possibilities. At IBM, Stanford professor Xiaochang Li has shown, a similar convergence of statistical knowledge, an engineering mindset, large sets of speech data, and access to computing power permitted a dramatic transformation of speech recognition.23 These industrial sites presaged—and made possible—much larger developments.
神经网络的地下世界
The Subterranean World of Neural Nets
尽管神经网络遭到了无数次打击，但从日本到法国，仍有一批研究人员继续研究神经网络，既是为了进行预测机器学习，也是为了更多地了解动物的大脑。尽管机器学习和人工智能社区中的许多人对神经网络持敌意态度，但贝尔实验室，尤其是加拿大组织 CIFAR 提供了必要的资金来维持研究，让人们记住网络的强大功能以及它们在识别数字等任务上取得的相对较新的成功。计算和神经科学交叉点的复杂故事在其他地方已经讲得很清楚了，所以我们只概述了关键的发展。24从非常普遍的层面上讲，我们可以说，到 20 世纪80年代中期，几位研究人员就如何训练多层神经网络提出了类似的想法，通过称为“反向传播”的过程。25当网络错误地将某些东西分类时，比如将热狗图像分类为狗，任何错误都会被用来改变网络深处的值，从而训练“神经元”犯更少的错误并做出更正确的决策。原则上，这种算法可以消除人工智能从业者拒绝神经网络的一些原因，因为这些“深度”网络可以区分比 20 世纪 60 年代的简单网络复杂得多的事物。并行计算机的发展使这项工作在计算上更合理。并行计算机涉及大量处理器来处理同一问题，而不是单个或少数几个单独工作的非常强大的处理器。
Despite the numerous strikes against neural nets, a passel of researchers from Japan to France continued to research neural networks, both for doing predictive machine learning and for learning more about animal brains. And, despite the hostility of many in the machine learning and AI communities to neural nets, Bell Labs and especially the Canadian organization CIFAR provided funding necessary to keep the research going, sustaining memories of the power of nets and their relatively recent successes at tasks like recognizing digits. The convoluted story at the intersection of computation and neuroscience is well told elsewhere, so we only sketch the key developments.24 At a very general level, we can say that by the mid-1980s, several researchers hit upon similar ideas for how to train multilayer neural nets, through a process known as “backpropagation.”25 When the network incorrectly classifies something, say classifying an image of a hot dog as a dog, any error is used to change values deeper into the network, thus training the “neurons” to make fewer errors and make more correct decisions. This algorithm could in principle put to rest some of the reasons for the rejection of neural networks at the hands of the AI practitioners, as these “deep” networks could discriminate among far more complex things than the simpler networks of the 1960s. The development of parallel computers made this work seem more computationally plausible. A parallel computer involves a large number of processors working on the same problem rather than a single or small number of very powerful processors working individually.
似乎旧符号人工智能人士的反对还不够，新数据密集型机器学习社区的许多人认为神经网络过时且浪费，是早期时代的倒退，早已被更好更便宜的算法所超越。与当时许多最好的算法不同，神经网络缺乏某些重要的数学特性，这让社区中的许多人感到沮丧。新的反向传播算法速度慢、计算密集，并且无法保证网络经过训练可以找到最佳答案——这是数学优化领域的核心标准，对早期人工智能人士和许多统计社区的人士都很重要。即使是新网络的倡导者也无法理解或详细解释网络为什么会做出这样的预测——它们是真正的黑匣子。它们在预测方面表现得足够好，但不是通过人类可以以任何普通方式理解的规则，而且计算成本非常高。这些技术在工业界取得了一些早期的成功，例如在读取银行支票数字方面，但直到 2010 年代，它们在学术界的声望都很低。
As if the opposition of the old symbolic AI folks wasn’t enough, many in the newly data-intensive machine learning community perceived neural networks as dated and wasteful, a throwback to earlier days, long since surpassed by better and cheaper algorithms. Unlike many of the best algorithms of the day, neural networks lacked certain important mathematical properties, to the dismay of many in the community. The new backpropagation algorithm was slow, computationally intensive, and provided no guarantee that the network had been trained to find the very best answer—a criterion central in the field of mathematical optimization and important to early generations of AI folks and many in more statistical communities. Even advocates of the new networks could not understand or explain in any detail why the networks made the predictions they did—they were truly black boxes. They worked well enough at prediction, but not through rules humans could understand in any ordinary way, and only at massive computational cost. The techniques found some early successes in industry, such as in the reading of numbers of bank checks, but carried little academic prestige well into the 2010s.
尽管新形式的神经网络取得了成功，但随后神经网络研究人员经历了一段近乎圣经般的流亡时期——至少在忠实信徒的眼中是如此。对于贝尔实验室的团队来说，反对者——他们许多最好的朋友——就在隔壁房间里，里面挤满了所谓的“内核”机器的支持者，当时这些机器似乎是不同机器学习算法之战的可能胜利者。26这些由 Vapnik 首创的内核机器具有强大的预测能力，但它们也具有重要的数学特性，深受数学倾向的研究人员的喜爱，而神经网络则缺乏这些特性。一位匿名的法国研究人员指出：“直到 2010 年，做网络都是‘过时的事情’。”其他研究人员对此兴趣不大，从对现任 Meta 首席人工智能科学家 Yann LeCun 的冷遇就可以看出。“我记得，LeCun 是实验室的受邀教授，我们不得不让某人和他一起吃饭。没人想去。” 27关键人物甚至离开了这个领域，尽管只是暂时的。尽管取得了明显的成功，但贝尔实验室招募的法国研究人员 Yann LeCun 和 Léon Bottou 却将注意力转向了创造更好的图像压缩方法。
Despite the successes of the new forms of neural networks, a nearly Biblical period of exile for neural net researchers followed—at least in the eyes of the true believers. For the team at Bell Labs, the opposition—many of their best friends—was in the next room, filled with proponents of so-called “kernel” machines that at the time seemed the likely victors in the battle among different machine learning algorithms.26 These kernel machines—pioneered by Vapnik—were powerfully predictive but they also had important mathematical qualities, beloved of mathematically inclined researchers, that neural nets lacked. “Until 2010,” one anonymous French researcher noted, doing nets was “a has been thing.” Other researchers had little interest, illustrated by the cold shoulder given Yann LeCun, now the chief AI scientist at Meta. “I remember, LeCun was in the lab as an invited professor and we had to make someone go to dinner with him. Nobody wanted to go.”27 Key people even left the field, albeit temporarily. Despite their apparent success, Yann LeCun and Léon Bottou for example, French researchers recruited to Bell Labs, turned their attention to creating better alternatives to compressing images.
然而讽刺的是，其他高度工具主义的预测算法的爆炸式成功为神经网络的普及铺平了道路。就在互联网数据和计算能力爆炸式增长的同时，算法系统的评估方式也发生了变化。到 20 世纪 90 年代，越来越多的文献表明，统计学家 Leo Breiman 认为，“将使用同一日期构建的多组预测因子结合起来，可以大幅降低测试误差。”然而，这种预测成功的代价是巨大的：这些模型对人类来说越来越难以理解。28这些令人惊叹的预测技术并没有让任何规则对人类来说变得可理解。然而，神经网络总是令人费解，而许多其他机器学习算法的一个基本优点是可解释。这些新的集成模型和神经网络的预测收益被广泛认为是巨大的，对于越来越多不同学科的从业者来说，这已经掩盖了所产生的预测集成的巨大不透明性。贝尔实验室的研究人员完成了大量庆祝集成建模的工作，并作为算法预测竞赛的获胜者参加，其中最著名的是 2009 年颁发的 Netflix 奖。越来越多的机器学习从业者放弃使用任何一种预测模型系列，转而组合许多不同的预测因子。集成的这种巨大成功放大了预测重于解释的道德。预测能力比任何其他优点都更能在机器学习中占据主导地位。对可解释性的要求、对人类可理解的规则的要求正在逐渐消退。
Ironically, however, the exploding success of other highly instrumentalist predictive algorithms was clearing the way for neural networks to become acceptable. The ground was shifting about how to evaluate algorithmic systems, just as data from the internet and computational power were exploding. By the 1990s, a growing literature revealed, the statistician Leo Breiman argued that “combining a multiple set of predictors, all constructed using the same date [sic], can lead to dramatic decreases in test error.” This predictive success came, however, at great cost: the models were increasingly inscrutable to human beings.28 These amazing predictive techniques did not render anything like rules intelligible to human beings. Neural nets, however, were always inscrutable, whereas a fundamental virtue of many other machine learning algorithms had been that they were interpretable. The predictive gains from these new ensemble models and neural nets were widely seen as massive, and to an increasing number of practitioners among different disciplines, have come to overshadow the massive opacity of the predictive ensemble generated. Bell Labs researchers produced much of the work celebrating ensemble modeling, and took part as the winners in contests of algorithmic prediction, most famously the Netflix Prize, awarded in 2009. Increasingly, machine learning practitioners were abandoning using any one family of predictive models in favor of combining many different predictors. This dramatic success of ensembles amplified the ethic of prediction over interpretation. Predictive capacity, more than any other virtue, became ever more ascendant in machine learning. Demands for interpretability, demands for rules understandable to human beings, were fading.
正是在这种背景下，神经网络才能从流放中回归，尤其是现在被重新命名为“深度学习”的大型多层神经网络。即使在 20 世纪 80 年代，神经网络的神秘性也使其存在问题，甚至令人怀疑；神经网络从 2012 年左右开始复兴，这完全取决于这种集成模型的合法性，无论是在商业、间谍还是科学领域。29
In precisely this context could neural networks return from their exile, in particular large, many-layer neural networks now rebranded as “deep learning.” Even in the 1980s, the inscrutability of neural nets made them problematic, if not suspect; the renaissance of neural networks from around 2012 rests squarely on the legitimation of such ensemble models, for commerce, for spies, and for science.29
2012 年，在预测 ImageNet 中物体的正确描述标签的年度竞赛中，一个神经网络以显著优势击败了所有其他竞争对手。ImageNet 是由斯坦福大学教授李飞飞及其团队收集的大型图像数据集。30到第二年，所有主要竞争对手都放弃了其他算法，转而采用自己的神经网络版本。31神经网络的支持者——长期以来一直被研究界排斥——感到自己得到了辩护。几十年来对他们方法的嘲笑突然变得显而易见。误导，记者和学者们开始讲述他们从不公正的流放中英勇归来的故事。
In 2012 a neural network dramatically outperformed all other contenders in an annual competition to predict the proper descriptive labels for objects found in ImageNet, a large data set of images assembled by Stanford professor Fei-Fei Li and her teams.30 By the next year, all the major competitors had abandoned other families of algorithms in favor of their own versions of neural nets.31 Supporters of neural nets—long shunned in the research community—felt vindicated. Decades of mockery of their approach suddenly seemed misguided, and journalists and academics alike began telling tales of their heroic return from unjustified exile.
整个竞赛依靠大量看不见的劳动力，通过亚马逊的 Mechanical Turk 进行，这是一个 2005 年推出的众包市场，任何人都可以雇佣大量远程工作人员来执行任务——通常是计算机还无法自动完成的任务。到目前为止，人工分类这些图像是亚马逊 Mechanical Turk 分布式劳动力的最大用途，因此也依赖于支付这些劳动力所需的大量资金，预计到 2010 年将包括约 25,000 人。32众包工作者将 1400 万张图片归入超过 21,000 个类别。他们在正确和错误分类方面的劳动为算法模型提供了“基本事实”，这些模型试图根据这些当时庞大的数据集进行预测。33
The entire competition rested on a vast abundance of unseen labor, through the use of Amazon’s Mechanical Turk, a crowdsourcing marketplace launched in 2005 that permits anyone to hire large numbers of remote workers to perform tasks—typically tasks computers can’t yet do automatically. The human classification of these images was the largest use of Amazon Mechanical Turk distributed labor up to that point, and thus also rested on the abundant funding needed to pay for that labor, estimated to include some 25,000 people by 2010.32 Crowdsourced workers put 14 million images into over 21,000 categories. Their labor in classifying, right and wrong, provided the “ground truth” for the algorithmic models to try to predict based on these enormous—for the time—data sets.33
神经网络在经历了漫长的沉寂之后，性能提升得益于大规模计算和海量数据集，因此其表现优于其他方法。神经网络在 2012 年的成功通常被描述为一次戏剧性的突破，但更为冷静的历史图景却表明并非如此。作为“深度学习”，神经网络在许多方面已被接受，因为其他模型已经远离了在短时间内计算的简单算法。神经网络面临的一个问题是训练成本高昂，因为它们需要大量的计算机时间和容量。到 2010 年，所有竞争对手的计算成本都相似，要么需要一台计算机进行长时间的训练，要么需要多台计算机并行工作——通常两者兼而有之。进行大量的计算工作需要大量资金。第二个问题是神经网络可能擅长（尽管启动缓慢）进行预测，但它们对这些预测的解释很少。但同样的情况也已成为现实竞争方法。到 2012 年，竞争对手使用极其复杂的算法组合，将它们捆绑在一起，形成一个整体，用于投票进行预测：它们几乎与神经网络一样复杂。与神经网络一样，大型集成模型、复杂的核空间和其他方法也同样将预测能力置于人类可解释性之上。
After their long time in the wilderness, with performance gains benefiting from massive computation along with such massive data sets, neural nets came to perform better than other approaches. Their success in 2012 is often portrayed as a dramatic break, but a more sober historical picture suggests otherwise. As “deep learning,” neural nets had become acceptable, in many ways because other models had moved far away from straightforward algorithms computed in short of amounts of time. One strike against neural nets was that they were expensive to train because they required so much computer time and capacity. By 2010, all the competitors were similarly computationally expensive, requiring either long training times with one computer or multiple computers working in parallel—usually both. Performing extensive computational work required large amounts of money. A second strike was that neural nets might be good at—if slow to start—making predictions, but they offered little explanation of those predictions. But the same had become true of competing approaches. By 2012, competitors were using extremely complex congeries of algorithms bundled together into an ensemble that voted in making predictions: they were almost as complex as the neural nets. Like neural nets, large ensemble models, complicated kernel spaces, and other approaches had similarly come to privilege predictive power over human interpretability.
直到哲学和数学上对算法系统几乎只关注预测的反对意见不再重要之后，深度学习才开始兴起——在工业界、军事界以及程度较小的学术界。人们普遍认为深度学习可以提供最好的预测器，因此，如果人们的目标是预测，那么深度学习就是最成功的方法。随着这些预测的成功，以及对算法系统应提供什么的期望缩小，统计学家和计算机科学家眼中神经网络的缺陷变得更容易被忽视。即使有了训练神经网络的新技术，这样做也需要大量的数据、巨大的计算能力和雄厚的资金来为大规模训练提供电力。
Deep learning ascended only after the philosophical and mathematical objections to algorithmic systems focused almost exclusively on prediction ceased to matter—in industry, in the military, and, to a lesser extent, in academia. Deep learning was widely understood to provide the best predictors, and thus the most successful approach if one’s goals were prediction. With these predictive successes, and the narrowing of expectations of what an algorithmic system should provide, the defects of neural nets in the eyes of statisticians and computer scientists became easier to ignore. Even with the new techniques for training neural networks, doing so required enormous amounts of data, huge computational power, and deep pockets to provide electricity for that training at scale.
人们对机器学习模型的期望发生了决定性的变化，而新的硬件变体，即图形处理单元 (GPU)，使神经网络的训练变得更容易、更快捷。34最重要的是，训练超大模型所需的现金越来越多地只通过谷歌和 GPU 制造商 NVIDIA 等公司提供给研究人员。此后的几年里，模型变得越来越大，在不断增加的数据集上进行训练，计算成本不断上升，无论是美元还是二氧化碳排放量。35
What people expected from machine learning models had decisively changed, and new variants of hardware, called graphics processing units (GPU), made training neural nets easier and faster.34 Above all, the cold hard cash needed to train extremely large models was becoming increasingly available only to researchers through companies like Google and the GPU manufacturer NVIDIA. In the years since, the models have only gotten larger, trained on ever-increasing data sets, with spiraling costs of computing, both in dollars and in carbon dioxide emissions.35
当神经网络开始流行时，机器学习的重新定义已经开始，重点是预测、大数据集和大型计算机。机器学习，尤其是使用神经网络的机器学习，被企业顾问和营销人员重新贴上人工智能的标签，有时这让研究人员感到不舒服。这类研究的规模和成本极大地改变了机器学习的学术研究甚至初创研究。只有少数公司拥有最前沿算法模型的数据、资金和计算能力，研究人员越来越依赖它们，如果不是为他们工作的话。AI Now 研究所教职主任、前谷歌员工 Meredith Whittaker 解释说：“想要开发和研究人工智能的大学实验室和初创企业发现，他们需要访问大型科技公司运营的昂贵的云计算环境，并争夺数据访问权，这种动态自 2012 年以来愈演愈烈。” 36她探讨道，用于机器学习的工具已经很容易获得并且越来越容易使用，但它们往往完全依赖于少数资源极其丰富的公司。 （例如，我们在哥伦比亚大学的课程利用了谷歌的产品 Colab，这让我们能够教授广泛的机器学习和统计技术，但代价是让学生适应使用谷歌工具。）
The redefinition of machine learning as focused on prediction, large data sets, and big computers was already under way when neural nets came into prominence. Machine learning, especially machine learning using neural nets, was rebranded as AI by corporate consultants and marketers, sometimes to the discomfort of researchers. The sheer scale and costs of this sort of research dramatically altered academic and even start-up research in machine learning. Only a few firms have the data, money, and computing power for the leading edge of algorithmic models, and researchers have come increasingly to depend on them, if not to work for them. “University labs and startups that wanted to develop and study AI found themselves,” AI Now Institute faculty director and ex-Googler Meredith Whittaker explains, “requiring access to costly cloud-compute environments operated by big tech firms and scrambling for access to data, a dynamic that has only intensified since 2012.”36 As she explores, the tools for doing machine learning have been accessible and increasingly easy to use, but often they depend utterly on a small number of extremely well-resourced firms. (Our course at Columbia, for example, makes use of Google’s product Colab, which allows us to teach a broad array of machine learning and statistical techniques at the cost of acclimatizing our students to the use of Google tools.)
为了什么而优化？
Optimizing for What?
2015 年，《科学》（科学界的《纽约客》）刊登了一篇文章，其中两位主要研究人员迈克尔·乔丹和汤姆·米切尔阐述了人工智能的发展状况。现在，人工智能的主导地位是从数据中学习模型，而不是通过硬编码规则。“许多人工智能系统开发人员现在认识到，对于许多应用而言，通过向系统展示所需输入输出行为的示例来训练系统要比通过预测所有可能输入的所需响应来手动编程系统容易得多。” 37这些算法的强大功能和适用性来自于要执行的任务范围的缩小。机器学习系统是否成功取决于用某种数字方式来表示对你来说重要的事情。作者解释说，“机器学习算法可以看作是在训练经验的指导下，在大量候选程序中搜索，以找到一个优化性能指标的程序。” 38换句话说，机器学习算法会产生大量候选程序来执行某项任务，比如对狗和猫进行分类，并根据你事先指定的指标寻找表现最好的程序：准确率，例如误报率最低。正如 Pat Langley 所抱怨的那样，“机器学习最初专注于使用和获取以丰富的关系结构形式呈现的知识，而现在许多研究人员似乎只关心统计数据。” 39
In 2015, Science—The New Yorker of the scientific world—carried a piece where two major researchers, Michael Jordan and Tom Mitchell, laid out the state of affairs in artificial intelligence. Having a model learned from data, rather than by hard coding of rules, now dominated AI. “Many developers of AI systems now recognize that, for many applications, it can be far easier to train a system by showing it examples of desired input-output behavior than to program it manually by anticipating the desired response for all possible inputs.”37 The power—and applicability—of these algorithms came from a narrowing of the tasks to be performed. A machine learning system is successful in terms of some numerical way of representing what matters to you. The authors explain, “machinelearning algorithms can be viewed as searching through a large space of candidate programs, guided by training experience, to find a program that optimizes the performance metric.”38 In other words, a machine learning algorithm produces a large number of candidate programs for doing some task, say classifying dogs vs. cats, and searches for one that best does so, according to a metric that you designate in advance: accuracy, with the smallest number of false positives, for example. As Pat Langley had complained, “Machine learning focused initially on using and acquiring knowledge cast as rich relational structures, while many researchers now appear to care only about statistics.”39
得到了什么？又失去了什么？预测占了上风——它胜过了对被预测事物的底层过程的建模。它胜过了对能够解释和理解算法在进行这些预测时的过程的担忧。神经网络长期以来一直令人厌恶，部分原因是它们不透明。但当大多数其他算法也变得同样不透明，并且基本目标是预测时，神经网络的缺陷就不再像以前那样重要了。40
What was gained? And what was lost? Prediction prevailed—prevailed over a modeling of the underlying processes of the thing being predicted. And it prevailed over a concern with being able to interpret and understand the processes of the algorithm in making those predictions. Neural networks were long anathema in part because they were opaque. But when most of the other algorithms had become similarly opaque, and the fundamental goals were predictive, the faults of neural networks no longer mattered as they once had.40
机器学习的这些巨大变化促成了企业界的蓬勃发展，并得到了其资金支持。这些指标至少间接地与金钱有关：页面浏览量、在线购买量、在社交网络上花费的时间、“参与度”。
These dramatic changes in machine learning enabled and were funded by its explosion within the corporate sector. There the metrics involve money, at least indirectly: page views, online purchases, time spent on a social network, “engagement.”
Netflix 奖
Netflix Prize
“我们真的很好奇，”Netflix 宣布，“金额为一百万美元。”2006 年，Netflix 提出了这个任何能够大幅改进向用户推荐电影的算法的人都将获得丰厚的奖金：
“We are quite curious, really,” Netflix announced, to “the tune of one million dollars.” In 2006, Netflix offered this substantial purse to anyone who could dramatically improve its algorithm for recommending movies to users:
我们想通过比赛来寻找答案。这真的很“简单”。我们为您提供了大量的匿名评级数据，以及一个比 [Netflix 的算法] 在相同训练数据集上所能达到的预测准确度高出 10% 的预测准确度标准。……如果您开发的系统在我们提供的合格测试集上被我们评判为大多数都超过了该标准，那么您将获得丰厚的奖金和吹嘘的权利。
we thought we’d make a contest out of finding the answer. It’s “easy” really. We provide you with a lot of anonymous rating data, and a prediction accuracy bar that is 10% better than what [Netflix’s algorithm] can do on the same training data set. . . . If you develop a system that we judge most beats that bar on the qualifying test set we provide, you get serious money and the bragging rights.
参赛者必须公开他们的算法：
Competitors had to make their algorithms public:
但是（你知道会有陷阱，对吧？）只有当你与我们分享你的方法并向世界描述你是如何做到的以及它为什么有效时，你才会这样做。41
But (and you knew there would be a catch, right?) only if you share your method with us and describe to the world how you did it and why it works.41
该公司提供了大量数据集：大约七年来对 17,770 部电影和 480,189 名匿名用户的评分，总计 100,480,507 条评分。虽然如此庞大的数据集正在提高大型互联网公司的知名度，但研究人员却很少能够访问这些数据集。贝尔实验室的 Chris Volinksy 解释说，Netflix“做出了一个绝妙的举动，他们意识到有一个研究社区正在研究这类模型，并且急需数据。” 42机器学习则不同，它更强大，拥有非常大的数据集。但它们仍然很少见。
The company made a substantial data set available: ratings on 17,770 movies and 480,189 anonymous users over approximately seven years, for a total of 100,480,507 ratings. While such large data sets were increasing the currency of large internet firms, researchers rarely had access. Bell Labs’ Chris Volinksy explained that Netflix “made a brilliant move by realizing that there was a research community out there that worked on these kinds of models and was starving for data.”42 Machine learning was different—more powerful— with very large data sets. But they remained rare.
2009 年，BellKor 的 Pragmatic Chaos 团队凭借构建出卓越的电影推荐系统赢得了 100 万美元奖金，以 20 分钟的优势击败了竞争对手 The Ensemble。获胜团队的名字集合了四个参与其中的独立团队的名字。
In 2009, the team BellKor’s Pragmatic Chaos won the million dollars for building a superior movie recommender system, beating out its competition, The Ensemble, by twenty minutes. The winning team’s name brings together the names of four separate groups who joined their efforts.
他们的社交能力体现在他们的获胜上算法将四组的努力整合成一个庞大的预测集合，将机器学习各个部分的模型整合在一起。由于缺乏可理解性或可解释性的约束，单一的性能指标使通过电子邮件和讨论板组织的特殊社交协调成为可能：一项竞争性的社区组织任务，即所谓的“共同任务框架”。数据科学家大卫·多诺霍 (David Donoho) 称，竞争重点是最大化共同得分，这是过去二十年机器学习在大型数据集上取得变革性成功的“秘诀”。共同任务允许“完全专注于优化经验性能，这……允许大量研究人员在任何给定的共同任务挑战中竞争，并允许对挑战获胜者进行有效和不带感情的评判。” 43多诺霍进一步指出，共同任务框架“立即导致现实世界中的应用。在赢得比赛的过程中，预测规则必然经过测试，因此基本上可以立即部署。” 44
Their social combination was mirrored in their winning algorithm, which combined the efforts of all four groups into a massive predictive ensemble, with models from all parts of machine learning brought together. Lacking constraints of intelligibility or explainability, the single performance metric enabled a peculiar social coordination organized through email and discussion boards: a competitive, community-organizing task, the so-called “common task framework.” The data scientist David Donoho calls the competitive focus on a common score to be maximized the “secret sauce” behind the transformative success of machine learning on large sets of data in the past twenty years. The common task allows “a total focus on optimization of empirical performance, which . . . allows large numbers of researchers to compete at any given common task challenge, and allows for efficient and unemotional judging of challenge winners.”43 The common task framework, Donoho argues further, “leads immediately to applications in a real-world application. In the process of winning a competition, a prediction rule has necessarily been tested, and so is essentially ready for immediate deployment.”44
事实上，部署机器学习通常涉及通过算法最大化量化值。在行业中，这种量化目标被称为“关键绩效指标”；这是一种与业务目标或产品目标（例如页面浏览量、在文章或视频上花费的时间或更普遍的“参与度”）相关的数值度量，或者理想情况下两者兼而有之！
Indeed, deploying machine learning often involves algorithmically maximizing a quantified value. In industry, such a quantitative goal is termed a “key performance indicator”; a numerical measure correlated either with business goals, or with product goals such as page views, time spent on an article or video, or “engagement” more generally—or ideally with both!
在 Netflix 竞赛结束时，麻省理工学院的研究员迈克尔·施拉格 (Michael Schrage) 解释说：“奖励模式的巨大优势在于，它让工作不再是选美比赛的范畴，而是以绩效为导向。”这样的庆祝自然建立在相信某些指标的优越性的基础上：“产生的结果才是最重要的。” 45说某件事重要或不重要是一种价值观的陈述。机器学习的支持者们并不看重美等复杂现象，而是更看重能够被量化的现象。在本书的开头，我们引用了德国人对新“粗俗”（定量）统计学家的不满，他们把数字误认为是有关土地或人民的知识，误解了价值。到 2000 年，机器学习发展起来，有望成为专注于数字的统计学家的典范——强大只是因为它的权限有限。当代人工智能的伦理和政治问题围绕着将人工智能重塑为指标的优化。
At the close of the Netflix competition, an MIT fellow, Michael Schrage, explained, “The great advantage of the prize model is that it moves work away from the realm of the beauty contest to being performance-oriented.” Such celebration naturally rests on believing in the superiority of some metric: “It’s the results produced that matters.”45 To say something matters or doesn’t is a statement of values. Rather than valuing complex phenomena like beauty, proponents of machine learning largely valued phenomena capable of being given a quantified measure. Early in this book, we quoted testy Germans upset at the new “vulgar” (quantitative) statisticians, who confused numbers for knowledge of a land or a people, and misunderstood value. Machine learning, as it developed by 2000, was poised to be an apotheosis of the numerically focused statistician—powerful just because it was limited in purview. The ethical and political concerns of contemporary AI circle around the recasting of AI as the optimization of metrics.
Netflix 竞赛表明，机器学习方法在 20 世纪 90 年代和 21 世纪的应用范围已远远超出学术中心和工业研究实验室，涉及商业、工业、医疗、警务和军事等一系列领域，这些领域既令人眼花缭乱，又令人兴奋，但又充满争议，有时甚至带有歧视性。到 2010 年代，那些提倡工业规模机器学习并将其融入商业和政府实践的人被称为“数据科学家”。在开发工具让从科学家到记者的每个人都能利用机器学习的同时，他们还利用了各种其他技能来扩展机器学习，使其在调解我们的通信、科学、新闻和政治的基础设施中发挥核心作用。
The Netflix competition illustrates how the machine learning approach came to be used in the 1990s and 2000s far beyond academic centers and industrial research laboratories, on an array of commercial, industrial, medical, policing, and military applications, at once dizzying, exciting, controversial, and sometimes discriminatory. Those advocating for industrial-scale machine learning and building it into business and governmental practices came to be known as “data scientists” by the 2010s. While producing tools enabling everyone from scientists to journalists to draw upon machine learning, they simultaneously drew on a broad array of other skills to scale machine learning, to make it figure centrally in the infrastructures mediating our communications, our science, our news, and our politics.
改变后的领域将被称为“数据科学”……数据科学的技术领域应该根据它们使分析师能够从数据中学习的程度来判断。
the altered field will be called “data science” . . . technical areas of data science should be judged by the extent to which they enable the analyst to learn from data.
–贝尔实验室统计学家 William Cleveland，《数据科学：扩展统计领域技术领域的行动计划》，2001 年
–Bell Labs statistician William Cleveland, “Data Science: An Action Plan for Expanding the Technical Areas of the Field of Statistics,” 2001
在 Facebook，我们觉得研究科学家、业务分析师等不同的职位名称不足以满足我们团队中各种工作需求。“数据科学家”可能会用 Python 构建多级处理管道、设计假设检验、使用 R 对数据样本进行回归分析、在 Hadoop 中设计和实现算法，或者以清晰简洁的方式向组织中的其他成员传达我们的分析结果。为了捕捉这些，我们想出了“数据科学家”这个头衔。
At Facebook we felt like different job titles like research scientist, business analyst didn’t quite cut it for the diversity of things that you might do in my group. A “data scientist” might build a multistage processing pipeline in Python, design a hypothesis test, perform a regression analysis over data samples with R, design and implement an algorithm in Hadoop, or communicate the results of our analyses to other members of the organization in a clear and concise fashion. To capture this, we came up with the title “data scientist.”
–Jeff Hammerbacher，《信息平台与数据科学家的崛起》，2009 年
–Jeff Hammerbacher, “Information Platforms and the Rise of the Data Scientist,” 2009
诗人艾伦·金斯堡写道： “我看到我们这一代最优秀的人才被疯狂摧毁。”金斯堡在一句又一句地歌颂着更高抱负与冷战时期美国现实之间的鸿沟：“天使般的嬉皮士们渴望与夜晚机器中的星光发电机建立古老的天堂联系”——以及学生在日益军事化的大学中所经历的鸿沟：“他们带着光芒四射的冷静走过大学战争学者们的眼睛里浮现出阿肯色州的幻觉和布莱克式的悲剧。” 1 2011 年，前 Facebook 数据团队负责人杰夫·哈默巴赫 (Jeff Hammerbacher) 在谈到金斯伯格时哀叹道：“我们这一代最优秀的人才都在思考如何让人们点击广告。这太糟糕了。” 2在所有需要优化的事情中，这一代人选择了操纵注意力。本章追溯了“数据科学”的演变过程，这个术语最初在让人们点击广告的公司中占据主导地位，但其历史从冷战一直延续到现在。
“I saw the best minds of my generation destroyed by madness,” wrote the poet Allen Ginsberg. In clause after clause, Ginsberg sang of the gulf between higher aspiration and the realities of Cold War America: “angelheaded hipsters burning for the ancient heavenly connection to the starry dynamo in the machinery of night”—and the chasm experienced by students with the increasingly militarized universities: “who passed through universities with radiant cool eyes hallucinating Arkansas and Blake-light tragedy among the scholars of war.”1 In 2011, Jeff Hammerbacher, a former Facebook data team leader, riffing on Ginsberg, bemoaned, “The best minds of my generation are thinking about how to make people click ads. That sucks.”2 Of all the things to optimize, a generation had chosen manipulating attention. This chapter traces the evolution of “data science,” a term which gained prominence first at companies making people click on ads, but whose history stretches from the Cold War to the present.
哈默巴赫与 DJ Patil 共同创造了“数据科学家”一词，用来描述从初创企业到财富 500 强企业等企业界的一个关键新角色。数据科学家所做的工作与我们所见的各种量化方法的实践者有何不同？什么是“数据科学”？我们会看到，它的定义各不相同。工业数据科学指的是机器学习和统计学，结合软件工程和具体的数据工作，以构建数字产品和服务。在学术研究中，这个术语含义广泛，不仅限于统计学，还包括通过数据理解世界所需的更广泛、不太“技术性”的技能，从“数据清洁工作”的混乱到通过数据传达结果的细微差别。这个术语不是抽象地“渴望古老的天堂联系”，而是谈到此类工作的实际复杂性，从数据分析开始，逐渐变得复杂。数据科学家乔尔·格鲁斯 (Joel Grus) 讽刺了另一位截然不同的冷战作家，他讽刺了人们对“数据科学家”掌握工业界所需的各种数据任务的期望：
Along with DJ Patil, Hammerbacher is credited with coining the term “data scientist” to describe a crucial new role in the corporate world from start-ups to Fortune 500 corporations. What does a data scientist do differently than practitioners of all the various quantitative approaches to the world we’ve seen? What exactly is “data science”? Definitions, we will see, vary. Industrial data science came to mean machine learning and statistics combined with the software engineering and concrete data work needed to build digital products and services. In academic research, the term is capacious, extending beyond statistics to include the broader and less “technical” skills needed for making sense of the world through data, from the messiness of “data janitorial work” to the nuances of communicating results through data. Rather than abstractly “burning for the ancient heavenly connection,” the term speaks to the practical complexities of such work, starting with data analysis getting grubby with data. Riffing on another, very different, Cold War writer, the data scientist Joel Grus satirized the expectation a “data scientist” had mastered the wide diversity of data tasks needed in industry:
数据科学家应该能够运行回归分析、编写 SQL 查询、抓取网站数据、设计实验、分解矩阵、使用数据框、假装理解深度学习、从 d3 图库中窃取、争论 r 与 python、在 mapreduce 中思考、更新先验、构建仪表板、清理混乱数据、测试假设、与商人交谈、编写 shell 脚本、在白板上编码、破解 p 值、机器学习模型。专业化适合工程师。*
a data scientist should be able to run a regression, write a sql query, scrape a web site, design an experiment, factor matrices, use a data frame, pretend to understand deep learning, steal from the d3 gallery, argue r versus python, think in mapreduce, update a prior, build a dashboard, clean up messy data, test a hypothesis, talk to a businessperson, script a shell, code on a whiteboard, hack a p-value, machine-learn a model. specialization is for engineers.*
随着该领域在业界和学术界声名鹊起，随之而来的工作机会、资金机会以及新部门和新学位的出现，雇主和管理人员试图更准确地定义它。通常，试图确定“数据科学”一词的含义会演变为与互联网共同发展的网络评论区中的口水战。我们不是坚持“数据科学”的一个定义，而是试图勾勒出围绕该术语的争论轮廓。十年来，在演讲、模因、帖子评论中，从业者一直在争论这个术语的真正含义，与统计学、机器学习或早期的“数据挖掘”形成鲜明对比。争论的根本问题是，谁有权力，谁有能力重新安排处理数据的权力。他们还关心谁最终会得到资金——来自企业、学术界和政府。
As the field rose to prominence in industry and academia, with associated job opportunities, funding opportunities, and new departments and degrees, employers and administrators sought to define things more precisely. Often, trying to nail down “data science” devolves into a verbal tussle in the online comment sections which coevolved with the internet. Rather than insist on one definition of “data science,” we seek to outline contours of contestation around the term. For a decade now, in presentations, through memes, in comments to posts, practitioners have fought over what the term really stands for, in contrast to say statistics, machine learning, or earlier “data mining.” The arguments fundamentally concern who has authority and who gains capacities to rearrange power in dealing with data. And they concern who ultimately gets the funding—in corporations, in academia, and from the government.
需要明确的是，人们对此感到兴奋并投入资金是有充分理由的。在各个行业中， 通过数据改变世界是革命性的。向商业用户推荐合适的产品和内容的能力使所谓的“长尾”商业模式成为可能。3同样，在商业软件方面，我们已经习惯了将手机作为我们可以“与之”交谈而不是“通过”的设备，因为语音识别已经通过多次量子飞跃得到改进。在金融领域，最赚钱的基金，文艺复兴科技的 Medallion Fund，使用统计分析进行交易，同时非常重视收集数据、学习模型和执行交易所需的软件工程。4在生物学和人类健康方面，人们很快意识到 20 世纪 90 年代的全基因组测序有可能通过数据改变我们对复杂人类疾病的理解。“生物学正处于一场智力和实验的巨变之中，”生物学家 Shirley Tilghman 在 2000 年《自然》杂志的一篇文章的第一句话中宣称。“从本质上讲，这门学科正在从一门数据贫乏的科学转变为一门数据丰富的科学。”在人类努力的各个领域，显然“新技术允许提出全新的问题”，“将需要……新的分析工具” 。5
To be clear, there was good reason for excitement and funding. In a variety of industries, making sense of the world through data had been transformational. The ability to recommend the right product and content to commercial users made possible a so-called “long tail” business model.3 Similarly, in commercial software, we’ve become used to phones as devices we can talk “to,” not “on,” as speech recognition has improved through multiple quantum leaps. In finance, the single most profitable fund, the Medallion Fund at Renaissance Technologies, trades using statistical analysis, along with considerable attention to the software engineering needed to gather data, learn models, and execute trades.4 In biology and human health, it was quickly realized that the sequencing of whole genomes in the 1990s had the potential to change our understanding of complex human diseases through data. “Biology is in the midst of an intellectual and experimental sea change,” declared the biologist Shirley Tilghman in the first sentence of an article in Nature in 2000. “Essentially the discipline is moving from being largely a data-poor science to becoming a data-rich science.” In a wide variety of fields of human endeavor, it was clear that “new technology permitted entirely new questions,” that “will require . . . new sets of analytical tools.”5
仅仅是统计数据还是不是...
Merely Statistics or Not . . .
2011 年，数学家、数据专家 Cathy O'Neil 和统计学家 Cosma Shalizi 在网上就当下“最性感”的职业——数据科学家的本质展开了一场善意的争论。O'Neil 认为，数据科学的大部分工作都涉及到使用统计数据的程度：
In 2011, a mathematician turned data guru, Cathy O’Neil, and a statistician, Cosma Shalizi, got into a good-natured tiff on the internet about the nature of the “sexiest” career of the moment: the data scientist. O’Neil argued that much of data science involves getting to the point one can use statistics:
换句话说，一旦我们把某件事归结为统计学问题，它就变得轻而易举了。即便如此，没有任何东西像你在统计课上实际发现的那样标准——被问到类似统计课的问题的可能性为零。
In other words, once we boil something down to a question in statistics it’s kind of a breeze. Even so, nothing is ever as standard as you would actually find in a stats class—the chances of being asked a question similar to a stats class is zero.
数据科学家对非标准化数据提出的问题范围更广。因此，他们需要不同的能力。
Data scientists had a much wider range of questions asked about less standardized data. And thus, they needed different competencies.
我想补充一点，定义数据科学家的并不是是否熟悉一套特定的工具。而是是否能熟练使用这些工具，成为一名工匠（和销售员）。
I would add that it’s really not about familiarity with a specific set of tools that defines a data scientist. Rather, it’s about being a craftsperson (and a salesman) with those tools.
打个比方：我并不是因为懂砂锅菜才当厨师的。6
To set up an analogy: I’m not a chef because I know about casserole dishes.6
Shalizi 表示反对：
Shalizi demurred:
但让我印象深刻的是，她所描述的优秀“数据科学家”所具备的技能只是优秀统计学家技能的一部分。它们最多只是优秀计算能力强的统计学家技能的一部分。7
What strikes me about it, though, is that the skills she’s describing a good “data scientist” as having are a subset of the skills of a good statistician. At most, they are a subset of the skills of a good computationally competent statistician.7
工业数据科学与统计学和机器学习等学术领域之间的主要区别在于，工业数据科学优先考虑并重视处理棘手的现实世界数据的能力，而这些数据通常来自大型基础设施。“处理”是指将数据转化为标准算法可以使用的形式。但它通常还意味着将非常大的数据集整理到能够处理这些数据的分布式数据库中。与这些学术领域不同的是，数据科学通常被理解为从根本上以组织（无论是公司还是政府机构）的业务需求为导向。
At the heart of the difference between industrial data science and the academic worlds of statistics and machine learning is the prioritizing and celebration of capabilities for dealing with problematic real-world data, often in large infrastructures. “Dealing with” stands for the craft of getting it into forms that standard algorithms can use. But it often also means wrangling very large data sets into distributed databases capable of contending with them. And unlike those academic fields data science is often understood as fundamentally oriented by the business needs of an organization, be it a corporation or government agency.
许多学术统计学家和机器学习者都认为这些更像手艺的元素在知识层面上较低，甚至可能简单易学。虽然理论性较低的科目不一定容易掌握，但还不够；正如奥尼尔打趣说的那样，砂锅菜的知识并不能成就大厨。
MANY ACADEMIC STATISTICIANS and machine learners have disparaged these more craft-like elements as lower on the knowledge scale and perhaps even simple to learn. Though less theoretical subjects are not necessarily easy to master, they are not enough; as O’Neil quipped, knowledge of casserole dishes does not a chef make.
最狂妄自大的说法是，数据科学被奉为一门大师级学科，能够重新定位科学、商业世界和治理本身。它是重组知识和权力的候选者，而这些知识和权力目前已融入主导我们大部分生活的机构之中。
At its most hubristic, data science is presented as a master discipline, capable of reorienting the sciences, the commercial world, and governance itself. It is a candidate for reorganizing knowledge and power, now built into institutions dominating much of our lives.
数据科学的根源是复杂的；它们既包括精湛的数学，也包括许多工程技术；既存在于大学演讲厅，也存在于营销部门和政客的作战室。数据科学的不纯粹混合性说明了我们一直在追踪的一个关键故事：日益自动化的决策形式与支持这些过程的大规模基础设施的结合。数据科学起源于统计学与现实世界数据的结合、机器学习以及大小型企业内数据分析处理的结合。这个故事需要在警告“过度数学化”的计算统计学家的世界和行业发展之间穿梭。我们从一些异端统计学家开始，他们根据自己的现实世界经验，劝告他们的学术领域更贴近数据。
The roots of data science are gnarled; they include rarified math but also much engineering craft; university lecture halls but also in marketing departments and the war rooms of politicians. The impure blended quality of data science speaks to a key narrative we have been tracing: the coming together of increasingly automatic forms of decision-making with large-scale infrastructures enabling those processes. Data science arises from a coming together of statistics working with real-world data, machine learning, and analytical processing of data within businesses large and small. The story requires moving between the world of computational statisticians warning about “over-mathematization” and developments within industry. We begin with some heretical statisticians who, informed by their real-world experiences, exhorted their academic field to move closer to the data.
“数据分析”，1960 年代至 1990 年代
“Data Analysis,” 1960s–1990s
1974 年，普林斯顿贝尔实验室的数学家 John Tukey 同意在美国国家安全局发表演讲，谈论“探索性数据分析”，要求该机构提供“2 个屏幕和 2 台投影仪，用于放映大型幻灯片”。8自从第二次世界大战期间参与密码学研究以来，图基一直担任美国国家安全局的科学顾问，自20世纪 40 年代以来，他一直在使用各种统计和图形方法创建新工具，用于探索大大小小的数据。最初，他专注于纸质数据探索工具，后来成为使用计算机绘制和分析数据的先驱。25 年前，美国国家安全局的库尔巴克邀请图基参加“数据存储和检索一般问题研讨会”——部分基于图基建议美国国家安全局调查该问题。9研讨会将讨论数据存储和检索的一般问题以及对美国国家安全局特别重要的问题。10
In 1974 the Princeton–Bell Labs mathematician John Tukey agreed to speak at the National Security Agency about “exploratory data analysis,” asking that the agency provide “2 screens and 2 projectors for large transparencies.”8 Long a scientific advisor to the NSA following his involvement with cryptography during World War II, Tukey had, since the 1940s, been creating new tools for exploring data, large and small, using all manner of statistical and graphical methods. Initially focused on paper tools for exploring data, he was at the forefront of the move to computers for graphing and analyzing data. Twenty-five years before, NSA’s Kullback had invited Tukey to a “symposium on the general problem of data storage and retrieval”—based in part on Tukey’s recommendation that the NSA look into the problem.9 The symposium was to consider what were the data storage and retrieval issues in general—and what were those particularly to the NSA.10
与图基所做的仍属机密的工作相比，他所提倡的对国家安全局和非机密世界内的统计数据和数据的态度并不重要。图基数十年来一直致力于将战争中大量数据集的实际统计工作转变为更为通用的工具集和思维方式。在他的职业生涯中，他涉足了从人口普查到导弹等各个领域。他所提倡的工具和他所提倡的图形技术（如箱线图）充斥着当代数据实践，包括中学标准化考试。
Less important than the still-classified work Tukey did were the attitudes toward statistics and data he encouraged within the NSA—and within the unclassified world. Tukey worked for decades to transform the practical statistical work on large data sets of the war into far more general use toolsets—and mindsets. In his career, he worked on everything from the census to missiles. The tools whose creation he encouraged and the graphical techniques he advocated such as the box-plot saturate contemporary data practices, including middle school standardized exams.
第二次世界大战期间，大规模数据分析成为必需，图基从中汲取了这一经验，提出了改变数据处理方法的纲领性声明，并试图制造工具来实现这一目标。在 1962 年的一份宣言中，图基呼吁采用一种他称之为“数据分析”的新方法，这种方法既致力于发现，也致力于确认：
Informed by the large-scale data analysis needed during World War II, Tukey provided a programmatic statement of a changed approach to data and sought to make tools to realize it. In a 1962 manifesto, Tukey called for a new approach he dubbed “data analysis” that would be dedicated as much to discovery as to confirmation:
数据分析以及与之相关的统计学部分必须具有以下特点：是一门科学，而不是数学，具体来说：
Data analysis, and the parts of statistics which adhere to it, must then take on the characteristics of a science rather than those of mathematics, specifically:
1.数据分析必须追求范围和实用性而不是安全性。
1. Data analysis must seek for scope and usefulness rather than security.
2.数据分析必须愿意适度地犯错，以便不充分的证据能够更频繁地暗示正确答案。
2. Data analysis must be willing to err moderately often in order that inadequate evidence shall more often suggest the right answer.
3. 数据分析必须以数学论证和数学结果作为判断依据，而不是作为证明依据或有效性标志。11
3. Data analysis must use mathematical argument and mathematical results as bases for judgment rather than as bases for proofs or stamps of validity.11
作为一种科学实践，图基将数据分析描述为一门艺术，而不是一门逻辑上封闭的学科。图基正在具体化一种替代学术统计学的方法，这种方法利用统计思维的数学能力进行探索和验证，并且可能适用于观察数据，而不仅限于作为实验试验的一部分产生的数据。正如我们上面所看到的，由于数学家米娜·里斯的支持以及统计学家哈罗德·霍特林等人的努力，第二次世界大战期间高度应用的统计学的巨大成功被引导为资金和象征性支持，以在美国和欧洲创建以数学为重点的理论统计学，而不是更注重实践、以数据为中心的统计学。不久之后，在图基等批评家眼中，实际的数据收集和分析就被牺牲在数学复杂性和严谨性的祭坛上。他反对大学统计学家的主流倾向，即强迫统计学尽可能模仿纯数学的抽象形式，在他看来，这种立场过于严谨，数据处理不足。回想一下，统计学家霍特林则担心年轻学生接触到的腐蚀性影响 实际数据太多。（需要说明的是，图基对数学严谨性并不陌生，他在第二次世界大战爆发的同一年完成了拓扑学博士学位，拓扑学是纯数学的一个分支。）
As a scientific practice, Tukey described data analysis as an art, not a logically closed discipline. Tukey was crystallizing an alternate approach to academic statistics, one that used the mathematical power of statistical thinking for exploratory as well as confirmatory purposes, and one that might be applicable to observational data, not exclusively to data produced as part of an experimental trial. Thanks to the support of the mathematician Mina Rees and the efforts of the statistician Harold Hotelling and others, as we saw above, the great successes of highly applied statistics during World War II were channeled into financial and symbolic support for the creation of a mathematically focused, theoretical statistics in the United States and in Europe, rather than a more practically oriented, data-focused statistics. Before long, in the eyes of critics such as Tukey, practical data collection and analysis had been sacrificed at the altar of mathematical sophistication and rigor. He was pushing against the dominant tendency of statisticians in universities to force statistics to emulate the abstract form of pure mathematics as much as possible, a position that, in his eyes, involved too much rigor and not enough working with data. Recall that the statistician Hotelling, in contrast, worried about the corrupting influence of young students being exposed to too much actual data. (To be clear, Tukey was no stranger to mathematical rigor, having completed his PhD in topology, a branch of pure mathematics, the same year World War II began.)
作为贝尔实验室的核心成员，图基利用自己在战争时期的经验以及数十年为美国国家安全局和军队工作的经验，从事数据分析工作。图基在一次采访中解释说，由于 20 世纪 40 年代的“战争问题”，“人们很自然地认为统计数据的目的是用于数据——也许不是直接用于数据，但最多是间接用于数据。现在，我无法相信其他有实践经验的人没有这种观点，但他们肯定——我想说——没有宣传它。” 12在 20 世纪 60 年代和 70 年代，图基和其他批评者抱怨说，在学术数理统计学及其相关分支（如计量经济学）中，很少有人将数据分析和判断形式的实践培养作为一项核心工作。正如我们在上一章中看到的，以模式识别形式出现的数据分析在其他地方蓬勃发展，在数理统计学和其他成熟学科的阴影下，在企业研究实验室和工程部门，以各种名称出现。
Tukey pursued this data analysis as a key member of Bell Labs, drawing on his wartime experience and decades-long work for the NSA and the military services. Thanks to “war problems” in the 1940s, Tukey explained in an interview, “it was natural to regard statistics as something that had the purpose of being used on data—maybe not directly, but at most at some remove. Now, I can’t believe that other people who had practical experience failed to have this view, but they certainly—I would say—failed to advertise it.”12 In the 1960s and 1970s, Tukey and other critics complained that relatively few within academic mathematical statistics and its allied branches, such as econometrics, celebrated the practical cultivation of data analysis and forms of judgment as a central endeavor. As we’ve seen in the previous chapter, data analysis in the form of pattern recognition flourished elsewhere, in the penumbra of mathematical statistics and other well-established disciplines, in corporate research labs and engineering departments, under various names.
在贝尔实验室的氛围中，图基和他的同事们创造了各种各样的统计和计算工具，使数据分析成为现实。十六年后，在一本实用教科书中，他解释说，“探索性数据分析”（EDA）是“侦探工作——数字侦探工作——或计数侦探工作——或图形侦探工作”。EDA 提供了一些适用于侦探工作各个领域的“一般理解”。 “刑事司法程序明确分为寻找证据——在盎格鲁-撒克逊国家，这是警察和其他调查部队的责任——和评估证据的强度——这是陪审团和法官的事。在数据分析中，类似的区分是有帮助的。探索性数据分析本质上是侦探性的。” 13探索性数据分析是一门技术手艺——而图基庆祝了为这门手艺创造的新工具。
In the atmosphere of Bell Labs, Tukey and his collaborators created a wide variety of statistical and computational tools needed to make data analysis a reality. Sixteen years later, in a practical textbook, he explained, “exploratory data analysis” (EDA) is “detective work—numerical detective work—or counting detective work—or graphical detective work.” EDA offered some “general understandings” useful across domains of detective work. “The processes of criminal justice are clearly divided between the search for the evidence—in Anglo-Saxon lands the responsibility of the police and other investigative forces—and the evaluation of the evidence’s strength—a matter for juries and judges. In data analysis a similar distinction is helpful. Exploratory data analysis is detective in character.”13 Exploratory data analysis is a technical craft—and Tukey celebrated the creation of new tools for that craft.
图基 1978 年出版的教科书草稿在贝尔实验室圈子及圈子以外流传多年，书中对通过有效的“重新表达”手段探索数据的艺术进行了概述。他用粗体字解释道：“在我们有效地展示结果之前，我们不会去看它们。 ” 14有效的展示意味着熟练掌握多种形式的数据可视化，图基强调，“需要付出更多的创造性努力来将数据分析的输出图形化……对于人类来说，使用合适的图片可以提供极大的灵活性，从广泛的总结到精细的细节，因为图片可以通过多种方式查看。”虽然图基预测计算机很快就会在绘图领域占据主导地位，但与此同时，他已经开发出各种手工可视化数据的做法。
Tukey’s 1978 textbook, whose draft had circulated for years in Bell Labs circles and beyond, offered a survey of the arts of exploring data through potent means of “reexpression.” “We have not,” he explained in bold type, “looked at our results until we have displayed them effectively.”14 Effective display means developing proficiency with many forms of visualizing data, Tukey emphasized, “much more creative effort is needed to pictorialize the output from data analysis. . . . For humans, the use of appropriate pictures offers the possibility of great flexibility all along the scale from broad summary to fine detail, since pictures can be viewed in so many ways.” While Tukey forecast that computers would soon predominate in graphing, in the meanwhile he had developed a variety of practices for visualizing data by hand.
贝尔实验室的同事们出色地沿着图基设定的路线继续工作，甚至超越了之前的路线，而现在，商业和科学系统中的数据爆炸式增长也日益成为大背景。1993 年，图基的贝尔实验室同事约翰·钱伯斯 (John Chambers) 撰写了自己的最新宣言，呼吁拓展统计学的抱负。钱伯斯抓住了只懂砂锅菜的人和厨师之间的区别，将“由文本、期刊和博士论文定义的”低级统计学与高级统计学进行了对比，高级统计学“包容性强，在方法论上不拘一格，与其他学科紧密相关，并且由学术界以外的许多人实践，通常也包括专业统计学以外的人”。15与低级统计学不同，高级统计学不仅关注简化、干净的数据，也不仅关注学术出版：
Colleagues at Bell Labs brilliantly continued work along the lines Tukey had set out and beyond, now increasingly in the context of the explosion of data in commercial and scientific systems. In 1993, his Bell Labs colleague John Chambers penned his own updated manifesto calling for the expansion of the ambitions of statistics. Capturing the difference between one who merely knows casserole dishes and a chef, Chambers contrasted lesser statistics, “as defined by texts, journals, and doctoral dissertations,” from greater statistics, “inclusive, eclectic with respect to methodology, closely associated with other disciplines, and practiced by many outside of academia and often outside professional statistics.”15 Unlike lesser statistics, greater statistics concerns itself not only with simplified, clean data, and not only with academic publication:
统计数字显示，工作主要有三大类：
Three broad categories characterize work in greater statistics:
—准备数据，包括规划、收集、组织和验证
— preparing data, including planning, collection, organization, and validation
—通过模型或其他总结来分析数据
— analyzing data, by models or other summaries
—以书面、图形或其他形式呈现数据16
— presenting data in written, graphical or other form16
钱伯斯坚持认为，在现实世界中准备和演讲“不仅具有实践意义，而且充满了智力挑战”。图基与美国国家安全局合作处理大量数据。钱伯斯指出了现实世界系统中数据不断积累所带来的挑战。“许多平凡的……活动会产生大量具有潜在价值的数据。例子……包括零售、结算和库存管理。这些数据不是为了学习而生成的，但是，学习的潜力是巨大的。” 17钱伯斯是在贝尔实验室的背景下写作的，可以访问该国的电信数据以及贝尔实验室研究人员在与美国政府合作时遇到的各种数据。
Preparing and presenting in real-world situations, Chambers insisted, was “rich with intellectual challenges as well as practical importance.” Tukey worked with the NSA to deal with vast data. Chambers noted the challenge presented by the increasing accumulation of data in real world systems. “Many mundane . . . activities generate large quantities of potentially valuable data. Examples . . . include retail sales, billing, and inventory management. The data were not generated for the purpose of learning; however, the potential for learning is great.”17 Chambers was writing in the context of Bell Labs, with access to the nation’s telecommunications data as well as the diverse data encountered by Bell Labs researchers in their partnerships with the US government.
类似的观察——为某一目的而收集的大型数据集可能会产生潜在的新型科学和商业知识——将在未来几十年内出现在各种计算领域。金融数据及其实际分析将催生技术分析、统计套利，后来随着计算工程的发展，高频交易领域也应运而生。同样，20 世纪 90 年代和 21 世纪的计算生物学也蓬勃发展，出现了对不同基因组的分析以及用于了解遗传网络的高通量生物检测、大规模电子健康记录挖掘和临床信息学。18在工业领域，应用计算统计方法改变了公司推荐书籍和电影的方式在电子商务兴起的早期，同样的技术就被应用于葡萄酒、鞋子，最终应用于信息和通信。每个领域都有自己的“数据时刻”，因为它重新发现了大量数据（这些数据不是用于学习）的价值，只要进行一些统计分析，并建立基础设施来收集、处理和产品化这些数据中的见解。钱伯斯、图基和其他人认为，统计分析只是这个项目的一部分——它是“更大”统计学核心的数学核心。但他们也警告说，如果学术统计学不开始提供从这些数据中学习的工具，它注定会变得无关紧要。
A similar observation—that large data sets gathered for one purpose may yield potential new kinds of scientific and commercial knowledge—would be made in a diversity of computational fields over the coming decades. Financial data and their practical analysis would give rise to technical analysis, statistical arbitrage, and later, with more computational engineering, the field of high frequency trading. Similarly, computational biology in the 1990s and 2000s exploded with analysis of differing genomes as well as high-throughput biological assays for understanding genetic networks, large-scale mining of electronic health records, and clinical informatics.18 In industry, applied, computational statistical methods changed the way companies recommended books and movies early in the rise of e-commerce, then later the same techniques would be applied to wine, shoes, and eventually information and communication. Each of these fields had its own “data moment” as it discovered anew how large quantities of data, generated for purposes other than learning, could be valuable given a bit of statistical analysis surrounded by an infrastructure need to gather, process, and productize insights from these data. Chambers, Tukey, and others argued that the statistical analysis was a mere part of this project—the mathematical nugget at the core of “greater” statistics. But they were also warning that academic statistics was doomed to irrelevance if it didn’t begin providing the tools for learning from this data.
1998 年，钱伯斯因用于数据分析和图形呈现的 S 系统获得了计算机协会的软件系统奖，“它永远改变了人们分析、可视化和处理数据的方式。” 19基于 S 的开源语言 R 成为计算导向统计学家的主要平台，特别是对于图形分析和呈现工作。
In 1998, Chambers received the Software System Award of the Association for Computing Machinery for the S system for data analysis and graphics presentation, “which has forever altered how people analyze, visualize, and manipulate data.”19 An open-source language R based on S became a dominant platform for computationally orientated statisticians and especially for work in graphical analysis and presentation.
因此，贝尔实验室的人们创造了工具集和态度，以实现数据分析，既包括传统统计形式，也包括更广泛的统计方法。他们还赞扬了图形方法的权力。
The Bell Labs crowd thus created tool sets and attitudes to enable data analysis, both of traditional statistical forms and of a broader approach to statistics. They also celebrated the power of graphical methods.
几年后，贝尔实验室的另一位统计学家威廉·克利夫兰明确呼吁创建“数据科学”领域，彻底改革统计学在实际数据分析中的实用性。统计学家可以为计算机科学家提供很多东西，而计算机科学家也可以教给统计学家很多东西：“计算机科学家对于如何思考和分析数据的知识是有限的，就像统计学家对计算环境的了解也是有限的一样。知识库的合并会产生强大的创新权力。这表明，统计学家今天应该从计算中获取知识，就像数据科学过去从数学中获取知识一样。” 20大学需要改变。
A few years later, another Bell Labs statistician, William Cleveland, explicitly called for creating the field of “data science,” a radical overhauling of statistics around its utility for practical data analysis. Statisticians had much to offer computer scientists—and computer scientists had much to teach statisticians: “the knowledge among computer scientists about how to think of and approach the analysis of data is limited, just as the knowledge of computing environments by statisticians is limited. A merger of the knowledge bases would produce a powerful force for innovation. This suggests that statisticians should look to computing for knowledge today, just as data science looked to mathematics in the past.”20 Universities needed to change.
贝尔实验室的异端统计学家——图基、钱伯斯、克利夫兰等人——并不是 20 世纪末唯一意识到将统计学应用于海量数据集可以创造一个新领域的人。工业界和学术界的人们也开始意识到这一点，他们试图创造存储、保护和搜索研究、商业和政府环境中日益产生的大型数据集所需的技术。到了 20 世纪 90 年代，许多从事所谓“超大型数据库”工作的社区人士担心缺乏能够分析日常在线和离线交易产生的数据的技术。21应对这一挑战需要新技术、新态度以及定义和授权一种新型从业者。
The heretical statisticians of Bell Labs—Tukey, Chambers, Cleveland, and others—were not the only ones in the late twentieth century to see that there was a new field to be created by applying statistics to massive data sets. This realization was also brewing among those in industry and in academia trying to create the technology needed to store, secure, and search the large data sets increasingly produced in research, commercial, and government settings. By the 1990s, many in the community working on what they called “very large databases” were worrying about the lack of technologies capable of analyzing the data produced through everyday online and offline transactions.21 Answering this challenge required new tech, new attitudes, and the definition and empowerment of a new kind of practitioner.
数据挖掘，20 世纪 90 年代初
Data Mining, Early 1990s
到 20 世纪 80 年代末，人们普遍认为，用于分析和学习快速增长的商业数据的工具越来越不充分，正如我们在第 8 章末尾看到的那样。科学、军事和情报数据也存在类似情况。1998 年，在大型企业、政府和学术“数据仓库”蓬勃发展的背景下，当时的微软研究员 Usama Fayyad 解释说：
By the late 1980s, the tools for analyzing and learning from the rapid expanding stores of business data were widely seen as increasingly inadequate, as we saw at the end of chapter 8. Similar stories held true with scientific, military, and intelligence data. In 1998, amid the blossoming of large-scale corporate, government, and academic “data warehouses,” then–Microsoft researcher Usama Fayyad explained:
如果我要用历史来类比我们今天在数字信息操纵、导航和利用方面的立场，我会想起古埃及……实际上，今天的大型数据存储与一个宏大的、只写的数据坟墓相差无几。22
If I were to draw on a historical analogy of where we stand today with regards to digital information manipulation, navigation, and exploitation, I find myself thinking of Ancient Egypt. . . . A large data store today, in practice, is not very far from being a grand, write-only, data tomb.22
许多有趣的大数据在两个方面是“大”的：它涉及对大量人员或大量购买的观察；对于每个观察，它涉及大量变量。最后一点——称为数据的“高维性”——伴随着重大的数学挑战。随着维度数的增加，用于比较数据点的数学技术变得困难重重，而获得更高可信度结论所需的数据量也变得更大。企业、军事和情报数据需要实时应对高维性的方法。
Much of the interesting big data is big in two different ways: it involves observations about say, a large number of people or a large number of purchases; and it involves, for each one of those observations, a large number of variables. The last point—called the “high-dimensionality” of data—has a major mathematical challenge that accompanies it. As the number of dimensions gets larger, the mathematical techniques used for comparing data points become problematic, and the amount of data necessary to achieve higher levels of confidence about conclusions becomes larger. Corporate, military, and intelligence data required the means for contending with high dimensionality in real time.
20 世纪 90 年代初，一场名为“数据挖掘”的运动兴起，旨在利用日益增长的未开发的企业和科学数据。数据挖掘，或更正式的名称是数据库中的知识发现 (KDD)，是从规模巨大、维度众多的数据库中创建适合采取行动的非凡知识的活动。23数据挖掘专注于非常大规模的数据库——数百万或数十亿条记录，通常每条记录都包含大量元素。对于零售数据库中的每条记录，数据挖掘操作可能会寻找所购买商品、商店邮政编码、购买者的邮政编码、信用卡种类、一天中的时间、出生日期、同时购买的其他商品、甚至查看的每件商品或之前购买或退回的每件商品的历史记录之间的意外关系。对高维、混乱的现实世界数据进行合理快速的分析是数据挖掘的本质和目的，甚至比模式识别或学术机器学习更为重要。 20 世纪 90 年代之前，复杂的统计和机器学习算法通常都是为那些可以轻松装入内存或只需要较少使用较慢磁盘访问的数据集而设计的。事实证明，将此类算法应用于无法装入内存的大量数据并非易事。 这不仅仅是将统计学或机器学习应用于更大的问题。数据挖掘的关键发展涉及努力在使算法可扩展所必需的权衡中进行选择。24反过来，应对规模的能力将极大地重塑实践中的机器学习。数据挖掘者大量借鉴了机器学习，他们为以新的方式扩展机器学习做了很多工作。从 20 世纪 80 年代末到现在，一种模式开始出现：算法倡导者采用一种特定的算法，提出一系列改进建议（通常作为博士研究的一部分），然后成为各个科学和工业领域中这些算法版本的倡导者。在潜在算法的狂野西部，一小部分从数据中学习的强大方法成为数据挖掘和机器学习中最有价值的算法。
A movement known as “data mining” emerged in the early 1990s to leverage the growing untapped stores of corporate and scientific data. Data mining, or, as it more formally was branded, Knowledge Discovery in Databases (KDD), is the activity of creating nontrivial knowledge suitable for action from databases of huge size and dimensionality.23 Data mining focused on databases of very large size—millions or billions of records, often with every record typically including a large number of elements. For each record in a retail database, a data mining operation might seek unexpected relationships among the item purchased, the store’s zip code, the purchaser’s zip code, variety of credit card, time of day, date of birth, other items purchased at the same time, even every item viewed, or the history of every previous item purchased or returned. Performing reasonably fast analyses of high-dimensional, messy real-world data is central to the identity and purpose of data mining, even more so than in pattern recognition or academic machine learning. Sophisticated statistical and machine learning algorithms before the 1990s were typically devised for sets of data that can easily fit in memory, or that require a relatively small use of slower disk access. Adapting such algorithms to huge quantities of data that cannot be held in memory proved far from obvious. It’s not just applying statistics or machine learning to a bigger problem. Key developments in data mining involve efforts to choose among the trade-offs necessary to make algorithms scale.24 The ability to contend with scale would, in turn, dramatically reshape machine learning in practice. Data miners drew heavily upon machine learning, and they did much to make it scale in new ways. In the late 1980s and up into the present a pattern begins to emerge: algorithmic advocates take up a particular algorithm, offer a series of suggested improvements, often as part of doctoral study, and then become advocates for versions of those algorithms across various scientific and industrial domains. From a wild west of potential algorithms, a small set of powerful approaches to learning from data emerged as the most prized algorithms in data mining and machine learning.
通过不拘泥于理论的数据，数据挖掘有望克服通常划分和理解世界的方式。按照费舍尔的模式工作的统计学家会问：“高收入人群是否比低收入人群更倾向于对仓储式商店忠诚？”并检验假设。“另一方面，数据挖掘可能会通过指出分析师原本无法考虑测试的、有助于商店忠诚度的其他因素，提供更多见解。” 25对这些方法的潜力感兴趣的科学家也提出了类似的说法。1999 年，帕特里克·布朗和大卫·博斯坦解释说：“探索意味着环顾四周、观察、描述和绘制未知领域，而不是检验理论或模型。目标是发现我们既不知道也不期望的事情。” 26
Through data unwedded to theory, data mining promised to overcome usual ways of dividing and understanding the world. Statisticians working in the mold of Fisher would ask, “Are higher-income people prone to be more loyal to a warehouse club than those with lower income levels?” and test the hypothesis. “Data mining, on the other hand, potentially would provide more insight by pointing out other factors contributing to store loyalty that the analyst would not otherwise have been able to consider testing.”25 Scientists intrigued by the potential of these approaches made similar claims. In 1999 Patrick Brown and David Botstein explained, “Exploration means looking around, observing, describing, and mapping undiscovered territory, not testing theories or models. The goal is to discover things we neither knew nor expected.”26
20 世纪 90 年代末，IBM 位于圣何塞的阿尔马登实验室举办了一系列研讨会，邀请了学术研究人员、工业研究人员和 IBM 自己的员工参加，并成为当地数据挖掘社区的社交中心。27 研讨会上提出的许多论文统计和机器学习算法的扩展将成为在现有硬件上使用大型数据集的标准变革性工作。
In the late 1990s, IBM’s Almaden Labs in San José hosted an ongoing seminar series that brought in academic researchers, industrial researchers, and IBM’s own employees, and more generally served as a center of sociability for the local data mining community.27 Many of the papers presented there would become standard transformative works in scaling statistical and machine learning algorithms for use in existing hardware with large data sets.
1997 年 11 月的一个星期三早上，一名斯坦福大学计算机科学研究生来到阿尔马登，就“网络挖掘”这一主题发表演讲。他解释说：
One Wednesday morning in November 1997, a Stanford computer science graduate student came down to Almaden to speak on the topic of “Mining the Web.” He explained:
斯坦福的一个新项目是 WebBase 项目。该项目的目标是从 Web 上收集大量数据，并将其用于研究。虽然该项目相对较新（几个月），但它已经产生了一些有趣的结果。
A new project at Stanford is the WebBase project. The goals are to collect a large amount of data from the Web and to make it available for research. While the project is relatively new (several months), it has already produced some interesting results.
演讲者谢尔盖·布林是数据挖掘小组 MIDAS（斯坦福数据挖掘）的组织者，该小组得到了几位教员的支持，他们都是数据库管理领域的先驱。在定期会议上，MIDAS 小组讨论了该领域的现状，从算法到道德：“话题范围从管理问题和拨款提案到学生和访客的会议式演讲。” 28在 IBM 的演讲中，布林承诺将广泛介绍他和斯坦福大学其他人为应对当时仍是新生事物的万维网的庞大规模而开展的工作。
The speaker, Sergey Brin, was the organizing force of a data mining group, MIDAS (Mining Data At Stanford), with the support of several faculty members, each a pioneer in database management. At its regular meeting, the MIDAS group discussed the state of the field, from algorithms to ethics: “Topics range from admistrative [sic] issues and grant proposals to conference-style presentations by students and visitors.”28 For his talk at IBM, Brin promised to range widely over work he and others at Stanford were doing to contend with the vastness of the then still novel World Wide Web.
我将谈谈我们利用这些数据发现的一些内容以及已经开发的一些算法，包括链接分析、质量过滤、搜索和短语检测。29
I will talk about some of the things we have discovered with this data and some algorithms that have been developed including link analysis, quality filtering, searching and phrase detection. 29
该项目很快便结出了许多算法成果。不久之后，该项目就获得了数十亿美元的收益。MIDAS 的网页指出，该团队“最令人印象深刻和最有用的演示”是“由拉里·佩奇和谢尔盖·布林创建的超级搜索引擎，名为 Google。” 30
The project would soon bear much algorithmic fruit. And, before long, many billions of dollars. The web page for MIDAS noted, “The most impressive and useful demo” of the group “is the super search engine, called Google, built by Larry Page and Sergey Brin.”30
口袋里的一张网能做什么：20 世纪 90 年代末
What Can You Do with a Web in Your Pocket: Late 1990s
20 世纪 90 年代，许多计算机科学界都发现自己没有做好充分准备来应对庞大、不断扩展且明显未经整理的万维网。旧式的搜索和索引工具是为高度标准化、整理、集中化的文本或其他数据集合（如期刊及其元数据集合）而设计的。研究人员在网页的非标准和非结构化质量及其数量方面苦苦挣扎。31与许多机器学习算法一样，信息检索这一古老领域的算法无法轻松扩展到网络页面数量。到 20 世纪 90 年代中期，搜索在许多人看来是一种没有前途的网络方法。主要的行业参与者越来越多地关注精心策划的门户网站，雅虎就是一个典型的例子。2000 年后，随着谷歌的逐渐崛起，搜索开始占据主导地位。他们的搜索方法正是源于对数据挖掘核心的高度应用机器学习的关注。
Numerous computer science communities found themselves underprepared in the 1990s to contend with the vast, expanding, and decidedly non-curated World Wide Web. Older search and indexing tools were designed for highly standardized, curated, centralized collections of text or other data such as a collection of periodicals with their metadata. Researchers struggled with both the nonstandard and unstructured quality of web pages and their number.31 Like many machine-learning algorithms, algorithms in the hoary field of information retrieval did not scale easily to the number of pages in the web. By the mid-1990s, search seemed to many an unpromising approach to the web. Major industry players focused increasingly on curated portals, exemplified by the approach of Yahoo. Search came to dominate after 2000, with the gradual, then exponential, rise of Google. And their approach to search emerged precisely from the concerns of the highly applied machine learning at the heart of data mining.
1998 年，斯坦福大学数据库组的 Brin 和人机交互组的研究生同学 Larry Page 提出了数据挖掘领域最著名的问题之一，即找出人们购物时哪些物品往往会搭配在一起，即所谓的“购物篮”问题。他们从大规模查看消费者购物篮中的物品中获取线索，寻找网络上文档之间的关联。他们的方法称为“动态数据挖掘”，并没有“详尽探索所有可能的关联规则空间”，因为网络太大而无法做到这一点：
In 1998, Brin, in the Database group at Stanford, and his fellow graduate student Larry Page, in the Human-Computer Interaction group, drew on one of the most famous problems of the data miners—figuring out what items tend to go together when someone shops—called the “market basket” problem. Taking cues from looking at items in consumers’ baskets at huge scale, they looked for associations within documents on the web. Their approach, called “dynamic data mining,” did not “exhaustively explore the space of all possible association rules”—as the web was far too big to do so:
当标准购物篮数据分析应用于购物篮以外的数据集时，在合理的时间内产生有用的输出非常困难。例如，考虑一个包含数千万个项目的数据大小，平均每个篮子有 200 个项目。……传统算法无法计算宇宙生命周期中的大型项目集。32
when standard market basket data analysis is applied to data sets other than market baskets, producing useful output in a reasonable amount of time is very difficult. For example, consider a data size with tens of millions of items and an average of 200 items per basket. . . . A traditional algorithm could not compute the large itemsets in the lifetime of the universe.32
正如机器学习算法必须改变以应对早期数据库挖掘的规模一样，关联挖掘算法也必须改变以应对早期万维网的规模。在将这种方法应用于商业数据库的过程中，布林和佩奇体现了专注于现实世界数据库的从业者在尽量减少磁盘和内存使用方面的动力。
Just as machine learning algorithms had to change to deal with the scale of early database mining, association mining algorithms had to change to deal with the scale of the early World Wide Web. In their adaptation of such an approach to commercial databases, Brin and Page exemplified the drive of practitioners focused on real-world databases to minimize disk and memory usage.
布林、佩奇和他们的其他合作者认为，互联网的规模使得它如此具有挑战性，但同时也使其前景广阔：
Brin and Page, along with their other collaborators, argued that the scale of the web, which made it so challenging, simultaneously made it deeply promising:
我们利用一个核心思想：网络提供自己的元数据。……这是因为网络的很大一部分是关于网络的。……由于网络的规模，专注于一小部分潜在有用数据的简单技术可以取得成功。33
we take advantage of one central idea: the Web provides its own metadata. . . . This is because a substantial portion of the Web is about the Web. . . simple techniques that focus on a small subset of the potentially useful data can succeed due to the scale of the web.33
Brin 和他的同事们身处一个对改造现有统计和机器学习技术有着浓厚兴趣的数据库社区，他们不仅准备应对规模问题，还准备将其打造成一个发现的中心资源。从根本上说，他们意识到网络的规模包括人类以数十亿种零碎的方式对网络进行分类和归类的巨大努力。他们没有创造任何能够通过编写规则对网络本身进行分类的人工智能，而是创造了一种大规模利用人类判断的机制。
Based within a database community deeply interested in transforming existing statistical and machine learning techniques, Brin and his collaborators were prepared not just to deal with scale, but to make it into a central resource for discovery. Fundamentally, they realized that the scale of the web included vast human effort to classify and categorize the web in billions of piecemeal ways. Rather than creating any form of artificial intelligence capable of classifying the web itself by writing rules, they created a mechanism for leveraging human judgment at great scale.
布林和佩奇在网络挖掘方面取得的最大突破是将日常学术实践转化为大规模算法形式，这种形式最有成效。他们根据佩奇的见解，采用了计算高质量引用次数的思想来衡量学术工作的权威性或价值。通过计算引用次数，即页面链接次数，可以对网页进行“排名”，以确定其权威性。更权威的页面是那些被其他权威页面链接的页面。指向某个页面的链接总数远远少于链接到该页面的页面的权威性。他们将结果称为 PageRank，并很快将其作为新搜索引擎 Google 的核心。Google 搜索诞生于一种文化中，这种文化融合了数据库的价值观，即扩展以处理数据和实践，后来又融合了机器学习社区的价值观。布林和佩奇从一开始就认识到需要构建能够在易犯错且功能有限的机器上实现精美数学运算的数据库。“Google 的数据结构经过优化，因此可以以较低的成本抓取、索引和搜索大量文档集。” 34 PageRank 是一种大规模利用人类判断力的过程，必须在一组经过创造性设计的数据库中实现。PageRank 及其在商品硬件中的实例化最终导致了分布式数据库和分布式分析处理的新架构的发展，分别称为 BigTable 和 MapReduce。正如我们将看到的，这些用于处理超大数据集的技术的发展在数据科学的后续发展中占据了核心地位。它们将先进的大规模机器学习变成了许多用户可以部署的技术——如果他们拥有合适的资源的话。
Brin and Page’s greatest breakthrough in mining the web came in adapting an everyday academic practice into algorithmic form most fruitful at vast scales. Following an insight of Page’s, they adapted the idea of counting high-quality citations to gauge the authority or value of academic work. Web pages could be “ranked” as more or less authoritative by counting citations, that is, links to pages. More authoritative pages are those that have been linked to by other authoritative pages. The total number of links to a page counted far less than the authority of the pages linking to that page. They called the result PageRank, and they soon made it central to a new search engine, Google. Google search emerged from within a culture which fused database values about scaling to deal with data and practice with, later, the values of the machine learning community. Brin and Page recognized from the start the need for structuring databases capable of implementing the beautiful mathematics on fallible and limited machines. “Google’s data structures are optimized so that a large document collection can be crawled, indexed, and searched with little cost.”34 A process for leveraging human judgment at mass scale, PageRank had to be materialized in a creatively designed set of databases. PageRank and its instantiation within commodity hardware in time led to the development of new architectures for distributed databases and distributed analytic processing, called BigTable and MapReduce respectively. The developments of these technologies for working with extremely large data sets figure centrally in the subsequent development of data science, as we will see. They made advanced machine learning at scale into technologies many users could deploy—if they had the right resources.
军事和情报方面的担忧从未远离这项工作的大部分内容。2004 年，国家安全局和海军研究办公室赞助了一次研讨会，讨论“海量数据流”的分析。9/11 事件发生后，情报和国防领域需要他们长期秘密培育的以数据为中心的企业的成果。数学研究小组的负责人解释了 NSA 从数据挖掘中获得了多少利润：
Military and intelligence concerns were never far from much of this work. In 2004, the National Security Agency and the Office of Naval Research sponsored a workshop on the analysis of “massive data streams.” In the wake of 9/11, the intelligence and defense worlds needed the fruits of the data-centric enterprises they had long secretly cultivated. The chief of the Mathematics Research Group explained how much the NSA was profiting from data mining:
我们在技术方面确实取得了一些显著的成功，这些技术是我们一年前没有的，比如在海量数据中寻找模式、得出结论、从某种情况中获取一些已知属性并通过数据挖掘来寻找新属性，这些技术非常基于算法，并且真正为我们的分析师提供了工具。……对我们来说，这一切都是为了教会机器如何为我们工作，而教会机器就是教会算法。35
We really have had some dramatic successes in terms of techniques we didn’t have a year ago for looking for patterns in massive data, drawing conelusions and taking some known attributes of a situation and mining through the data to find new ones, and very algorithmic based, and really providing tools for our analysts. . . . For us, it is all about teaching the machines how to work for us, and teaching the machines is teaching the algorithms.35
在 21 世纪，美国国家安全局并不是唯一一家从这种以数据为中心的计算工作中获利的机构；它很快被引入到营销、医学、物理、教育、刑事判决、社交网络和无人机瞄准等领域。在商业、情报和军事数据领域扩展机器学习所需的挑战促进了能够处理越来越大的数据集的技术和技术人员的创造。
In the 2000s, the NSA was not alone in profiting from this data-focused computational work; it was quickly imported into marketing, medicine, physics, education, criminal sentencing, social networking, and drone targeting. The challenges required to scale machine learning in the arenas of commercial, intelligence, and military data promoted the creation of technologies and technologists capable of dealing with ever-larger data sets.
从数据挖掘到大数据，2000-2010
From Data Mining to Big Data, 2000–2010
尽管“数据科学家”一词以前也曾被使用过，但当它作为职位出现在互联网平台 Facebook 和 LinkedIn 上时，它便开始流行起来。这些公司与学术界内部的斗争相去甚远，就像它们的竞争对手谷歌和亚马逊一样，它们正在以越来越快的速度从网上和网下的日常交易中积累数据，可能只有美国国家安全局本身能与之匹敌。存储、呈现、分析如此庞大的数据量需要巨大的技术和智力挑战，这些挑战与在台式计算机上分析较小的数据集在规模上截然不同。一旦解决了规模挑战，统计学家的技能、实践和软件才会在很久以后才需要。
Although used from time to time before, the term “data scientist” flourished when it appeared as a job title at the internet platforms Facebook and LinkedIn. Far distant from internecine academic battles, these firms, like their rivals Google and Amazon, were accumulating data from everyday transactions on and off the web at an ever-increasing rate, probably rivaled only by the NSA itself. Storing, presenting, and analyzing this tremendous volume of data entailed staggering technical and intellectual challenges, challenges radically different in scale from analyzing smaller data sets on a desktop computer. The skills, practices, and software of statisticians would only be needed far later, once the challenge of scale had been met.
随着互联网公司纷纷记录尽可能多的用户数据，数据也在快速积累。随着广告模式在此类信息平台公司中越来越受欢迎，公司客户的数据也越来越多。大约在同一时间，美国国家安全局获得了新的授权，允许其捕获大量互联网和电话流量，这超出了其分析能力。数据库和分析它们的能力正在崩溃。软件和硬件一次又一次无法处理这些流量。
Data was accumulating fast as a succession of internet companies recorded as much as they could about their users, and, with the rise in the prominence of the advertising model among such information platform companies, about their corporate clients. Around the same time, the NSA received new authorities allowing it to capture untold amounts of internet and telephony traffic that overwhelmed its analytic capacities. Databases and the ability to analyze them were collapsing. Time and again the software and hardware couldn’t handle the stream.
例如，当 Facebook 的一个关键数据库的数据量达到 1TB 时，Jeff Hammerbacher 解释道，查询系统“突然停止运行”。需要三天时间才能恢复。最终，Facebook 采用了 Hadoop，这是一个强大的开源框架，用于存储和分析大量数据；该技术主要由雅虎开发，允许将数据存储在数百台服务器上，并允许基于 Google 的 MapReduce 流程在这些服务器之间进行分析。Hadoop 还允许混合使用“结构化”和“非结构化数据”——想象一下地址，有明确划分的区块（姓名、街道、邮政编码），而不是一封信中连续的文本流。
For example, when a key Facebook database approached a terabyte of data, Jeff Hammerbacher explained, the querying system “came to a sudden halt.” It took three days for it to come back. Eventually Facebook adopted Hadoop, a powerful open-source framework for storing and analyzing large amounts of data; in large part developed by Yahoo, this technology allowed data to be stored over hundreds of servers and allowed analysis, based on a Google process known as MapReduce, to be divided among those many servers. Hadoop also allowed for a mix of “structured” and “unstructured data”—think of an address, with clearly demarcated blocks (name, street, zip code) versus the unbroken flow of text of a letter.
类似的故事在新旧行业中都有所体现，在学术领域中，新的数据宝库（尤其是通用数据）压倒了旧的计算分析模式。
Similar stories ramified across industries old and new— and within academic precincts where new troves of data, especially generic data, overwhelmed old modes of computational analysis.
如果数据太多是个问题，它还提供了很好的机会。三位谷歌研究人员对他们所谓的“数据的不合理有效性”大加赞赏。他们认为，简单模型下的大量数据几乎总是比数据较少的复杂模型效果更好。36 Facebook和谷歌努力利用这种新方法——美国国家安全局也是如此。
If too much data was the problem, it also offered great opportunity. Three Google researchers celebrated what they called the “unreasonable effectiveness of data.” They argued tons of data with simple models would almost always do better than complex models with little data.36 Facebook and Google worked to leverage this new approach—as did the NSA.
1996 年，美国国家安全局在一份高度机密的内部杂志上进行了一次采访，探讨了被监视的全球通信量的问题：
A 1996 interview in the highly classified house magazine of the NSA turned to the question of the volume of world communications to be spied upon:
让我补充一下我们面临的第三大挑战，那就是数量。我就可以结束这句话了，一切都说完了。37
Let me add to all of that the third biggest challenge facing us, and that is volume. And I could just end the sentence there and everything is said.37
2006 年，一封绝密邮件“数据量是我们的朋友”表明，人们对 NSA 应对数据过载的能力有了新的信心：事实上，数据量是其他地方推崇大数据的关键。数据量越大越好。
By 2006, a top-secret email “Volume is our Friend” suggests a newfound confidence in the NSA’s ability to contend with data overload: indeed, the enabling quality is central to the celebration of big data elsewhere. The bigger the volume, the better.
用于全球反恐战争的资源使 NSA 获得了巨大的收集和分析能力——但它并不需要更多。2008 年，NSA 转向使用开源社区内部制作的、基于谷歌创意的大数据数据库。基于谷歌的一些核心创意 BigTable，NSA 内部的一组科学家和程序员创建了一个分布式数据库平台——并发布给开源社区——用于容纳需要 PB 级存储容量的数十亿点的图表。38反恐战争期间，美国联邦政府向机器学习、计算统计和分布式计算投入了数百万甚至数十亿美元的资金。情报机构和军事部门大量借鉴学术和商业发展成果，同时为所有关键领域提供稳定的资金流。
The resources dedicated to the Global War on Terror allowed the NSA to acquire enormous collection and analytical capacity—but it ever needed more. In 2008, NSA turned to the databases for big data produced within the open-source community and based on ideas from Google. Based on some ideas central to Google called BigTable, a group of scientists and programming within the NSA created—and released to the open-source community—a distributed database platform designed to accommodate graphs with billions of points requiring petabytes of storage capacity.38 The War on Terror saw millions if not billions of US federal dollars pour into machine learning, computational statistics, and distributed computing. The intelligence agencies and military branches drew heavily upon academic and commercial developments, while providing a consistent stream of funding in all of the key fields.
2013 年左右，美国国家安全局发布招聘信息，要求招聘 SIGINT“信息专家”：
An NSA job posting from around 2013 calls for a SIGINT “Informatist”:
混合型计算机科学家和分析师，其工作跨越了以过程为中心的技术工作和以内容为中心的分析工作之间的界限。
hybrid computer scientist, analyst with work spanning the divide between the process-focused technical work and the content-focused analytical work.
职责：
Responsibilities:
• 将有关数据结构、语法和处理的信息与收集、组织和操作数据集的功能相结合，以综合满足客户信息需求的响应。
• Combine information about the structure, syntax, and processing of data with the functions of gathering, organizing, and manipulating datasets in order to synthesize responses to customer information needs.
• 应用科学技术进行数据评估、执行统计推断和数据挖掘。
• Apply scientific techniques to data evaluation, performing statistical inference and data mining.
• 记录并呈现数据分析及其结论，以供全绩效分析师、开发人员及其经理评估。39
• Document and present the data analysis and its conclusions for assessment by full-performance analysts, developers, and their managers.39
无论是在秘密世界还是在企业界，数据科学家这一新角色都变得越来越重要。与工业界和学术界一样，国安局也接受了企业所推崇和赋予的卓越和知识形式的转变。简要概述国安局从冷战到现在的机构文化的诸多变化，就可以解释该机构如何放弃“完美主义”文化，转而追求一种截然不同的文化。在冷战后期的“胜利”中，“国安局在 20 世纪 80 年代看重的是准确性、渊博的知识、全面的专业技能、生产力和声誉”。相比之下，在众多潜在敌人的不对称世界中，“国安局在 21 世纪看重的是……速度——现在掌握 80% 的速度可能会对挽救生命产生重大影响。”（当然，如果目标是信息，那么 20% 的时间里就意味着杀害无辜者。）” 40分析既深刻又可怕。Netflix 做出糟糕的建议是一回事，为监视、无人机袭击或更糟糕的情况提供依据则是另一回事。
In the secret world and in the corporate world alike, the new role of data scientist came to ever greater prominence. As in industry and academia, the NSA embraced a shift in the forms of excellence and knowledge that enterprises celebrated and empowered. A short overview of the many changes in the NSA’s institutional culture from the Cold War to the present explained how the agency had abandoned a culture of “perfectionism” for something radically different. Amid the “winning” of the late Cold War, “NSA Valued in the 1980s, Accuracy, Deep Knowledge, Thorough Expertise, Productivity and Reputation.” In the asymmetric world of a dizzying array of potential enemies, in contrast, “NSA valued in the 2000s . . . Speed—getting it 80 percent right now could make all the difference in saving lives. (Of course, if it were targeting information that would mean killing innocents 20 percent of the time.)”40 The analysis is incisive and scary. It’s one thing for Netflix to make bad recommendations, quite another to advance grounds for surveillance, drone strikes, or worse.
人工智能
Artificial Artificial Intelligence
凯瑟琳·迪伊格纳齐奥和劳伦·克莱恩在《数据女权主义》一书中坚持一个关键原则：“数据科学的工作，就像世界上所有的工作一样，是许多人共同努力的结果。” 41虽然互联网使数据收集变得前所未有的简单，但它并没有使数据处理独立于人类。大规模算法系统并没有消除人类的劳动和判断，而是取代了劳动力，并从根本上依赖于其他形式的劳动。所有新硬件和软件、所有算法的背后都是人类的工作，使数据变得易于处理。其中一些劳动属于新晋数据科学家的职责范围，但大部分（如果不是大多数的话）工作都落在了公司运作中很少见的人身上。计算机接管所有工作的愿景是错误的。“为了理解自动化对人类活动的影响，”学者安东尼奥·卡西利坚持认为，“我们必须首先认识和估计自动化本身的工作量。” 42
In their Data Feminism, Catherine D’Ignazio and Lauren Klein insist on a key principle: “The work of data science, like all work in the world, is the work of many hands.”41 While the internet makes the collection of data easy at hitherto unknown scales, it didn’t make processing it independent of human beings. Rather than eliminating human labor and judgment, large-scale algorithm systems both displace labor and fundamentally depend on other forms of labor. Underlying all the new hardware and software, all the algorithms, was human work to make the data tractable. Some of this labor fell within the purview of shiny new data scientists, but much if not most of the mucking fell to people rarely visible in the workings of the corporations. Visions of computers taking over all jobs are just false. “In order to understand what automation does to human activity,” insists the scholar Antonio Casilli, “we must recognize and estimate first the amount of work inscribed into automation itself.”42
数据劳动并非新鲜事物，在我们的历史上也并非如此：想想布莱切利园或十九世纪末的人口普查工作者。然而，今天的规模无疑是史无前例的——而且正是由相关系统促成的。
Data labor is not new, and nor is its obfuscation in our history: think of Bletchley Park or census workers in the late nineteenth century. Yet the scale today is certainly unprecedented—and enabled by the very systems in question.
从一开始，谷歌的搜索算法就利用了网页的隐性人工排名。为了获得该公司现在提供的相对干净的搜索结果，它依靠数十亿人为的判断来判断内容是否露骨、是否性别歧视、是否种族歧视，正如 Sarah Roberts、Mary L. Gray 所说，Siddharth Suri 和 Antonio Casilli 通过详细的人类学实地考察和社会学研究进行了记录。43在Lucy Suchman 和 Shoshana Zuboff 早期研究的基础上，这些学者都强调了从印度和菲律宾到美国农村地区，世界各地对劳动者的混淆过程，这种混淆让人觉得是技术在做工作，而不是人。Gray 和 Suri 解释道：
From the start, Google’s search algorithm leveraged the implicit human ranking of web pages. To get the relatively clean search results the company now delivers, it rests on billions of human judgments to deem content explicit or not, sexist or not, racist or not, as Sarah Roberts, Mary L. Gray, Siddharth Suri, and Antonio Casilli have documented through detailed anthropological fieldwork and sociological study.43 Building upon earlier work by Lucy Suchman and Shoshana Zuboff, these scholars all stress the processes of obfuscation of laborers worldwide, from India and the Philippines to the rural United States, an obfuscation that makes it appear the tech is doing the work—not the people. Gray and Suri explain:
每天，数十亿人都在消费网站内容、搜索引擎查询、推文、帖子和移动应用服务。他们认为，他们的购物完全是靠技术的魔力实现的。但实际上，他们是由一群在后台默默工作的国际员工提供服务的。44
Billions of people consume website content, search engine queries, tweets, posts, and mobile-app-enabled services every day. They assume that their purchases are made possible by the magic of technology alone. But, in reality, they are being served by an international staff, quietly laboring in the background.44
这样的劳动使得统计学和机器学习应用于大型数据集成为可能。卡西利解释道：“在投资者和媒体名人的想象中，机器人幻想的对立面是无数非专业的点击工作者，他们执行选择、改进和解释数据的必要工作。” 45将机器学习应用于世界需要数据，即使是自动收集的数据，也需要变得可用。
Such labor makes the application of statistics and machine learning to large data sets possible. Casilli explains it: “At the antipodes of robotic fantasies sustaining the imagination of investors and media personalities are the myriad nonspecialized click-workers performing the necessary work for choosing, improving, and making data interpretable.”45 Applying machine learning to the world requires data, even automatically collected data, to be made usable.
批评者正确地指出，许多所谓的人工智能成功在某种程度上涉及继续进行人类决策，而且往往是大规模的决策。教授兼前谷歌员工莉莉·伊拉尼 (Lilly Irani) 认为，亚马逊的 Mechanical Turk “通过用真人模拟人工智能对计算智能的承诺，让典型的人工智能项目得以推进。” 46但即使是自动化程度更高的系统也依赖于由劳动团队分类、清理和生成的数据，这些劳动团队通常远离软件公司的娱乐环境，他们的私人厨师和桌上足球。即使系统在执行类似人类的任务方面变得更优秀，它们通常也是基于更大的人类分类和生成的数据池。“这些工人推动着科技行业，”伊拉尼进一步解释道，“但在媒体和科技工作场所多元化政策中，他们却被忽视了。多样性是存在的。只是分包了，报酬很低。”她继续说：“这些工人擅长做机器做不到的事情。他们赢得了与机器的竞赛，但他们甚至连最低工资都拿不到。” 47这对数据科学来说也是如此，有好有坏。
Critics have rightly noted that many supposed AI successes at some level involve continuing human decisionmaking, often at a vast scale. Amazon’s Mechanical Turk, professor and former Googler Lilly Irani argues, “has allowed canonical AI projects to proceed by simulating Al’s promise of computational intelligence with actual people.”46 But even far more automated systems depend on data classified and cleaned and produced by teams of laborers, usually far from the ludic environs of the software companies with their private chefs and foosball tables. Even as systems became superior at performing human-like tasks, they usually do so based on larger pools of human classified and produced data. “These workers power the tech industry,” Irani explains further, “yet are out of sight and out of mind in the press and policy on diversifying the tech workplace. The diversity is there. It’s just subcontracted and paid poorly.” She continues: “These workers excel in doing what machines cannot. They have won the race against the machine, but they do not always even make minimum wage.”47 And this is true for data science, for good—and for bad.
统计学进入数据科学
Statistics Comes to Data Science
2014 年，伯克利大学教授于斌在悉尼举行的国际统计学家大会上发表主席演讲时提出：“让我们拥有数据科学。” 48 2010 年代初，记者、咨询公司和思想影响者都在庆祝数据科学家成为十年来最性感的职位。然而，最接近理解数据的学术领域——统计学——似乎已被抛在一边，过时了，被认为是错误的方法。她说，统计学家需要更多地参与计算、当代大数据形式和通信实践。
At a presidential address before an international conference of statisticians in Sydney, Berkeley professor Bin Yu proposed in 2014, “Let us own data science.”48 In the early 2010s, journalists, consulting firms, and thinkfluencers were celebrating data scientists at the sexiest job title of the decade. And yet the very academic field closest with understanding data— statistics—seems to have been left in the dust, old-fashioned, perceived as the wrong approach. Statisticians, she said, needed to become more engaged in computing, in contemporary forms of large data, and practices of communications.
在阐述自己的观点时，余解释说：“我们许多有远见的统计学同事都预见到了数据科学的到来。”余说得没错：统计学家有着悠久的传统，他们专注于数据、计算潜力和现实世界的应用。但他们大多逆流而上，反对我们所遇到的统计学的激进数学化，这种反经验主义精神在符号人工智能中占主导地位。这些叛逆的统计学家往往拥有双重国籍，既有学术界的，也有学术界以外的，通常在工业界和政府资助的研究中心工作。
In making her case, Yu explained, “Many of our visionary statistics colleagues saw data science coming.” Yu was not wrong: a rich tradition of statisticians had focused on data, the potential of computation, and real-world applications. But they had largely swum upstream, against the aggressive mathematization of statistics we encountered, the same antiempiricist spirit that dominated in symbolic AI. These renegade statisticians tended to have dual citizenship, within the academy and without, usually in industry and government-sponsored research centers.
除了 John Tukey，没有人比 Leo Breiman 更能体现叛逆的统计学家了，他从学术界转战到工业界。当他从工业和国防工作转回加州大学伯克利分校的学术界时，他大吃一惊。他后来形容这就像身处爱丽丝梦游仙境。
Alongside John Tukey, no one better embodied renegade statisticians than Leo Breiman, who shuttled from academia to industry. Upon moving from industry and defense work back into academia at UC Berkeley, he was startled. He later described it as being in Alice in Wonderland.
我知道工业界和政府在统计应用方面正在发生什么，但学术研究方面发生的事情似乎离我很远。它的发展就像是抽象数学的某个分支。49
I knew what was going on out in industry and government in terms of uses of statistics, but what was going on in academic research seemed light years away. It was proceeding as though it were some branch of abstract mathematics.49
放弃加州大学洛杉矶分校的数理统计学专业后，他为国防部和当时新成立的环境保护局从事了广泛的统计工作。他解释说，除了学术统计之外，他还研究污染和追踪苏联潜艇等课题，他开始专注于预测，而不是使用模型做出因果断言或进行严格的假设检验。50在学术界之外，布雷曼经历了——或者说是巩固了——他的认识论价值观和数学实践的根本性转变，从解释转向预测。
Having left a promising career in mathematical statistics at UCLA, he took on a wide range of statistical work for the Department of Defense and the then-new Environmental Protection Agency. Working outside of academic statistics on subjects such as pollution and tracking Soviet submarines, he explained, he came to focus on prediction over making causal claims using models or doing rigorous hypothesis testing.50 Outside of academia, Breiman underwent—or perhaps cemented—a fundamental shift in his epistemic values and mathematical practices, away from explanation to prediction.
统计学诞生于对不同人群和系统的数据进行分析，然而在 Breiman 这样的从业者眼中，这门学科已经误入歧途；直到 2000 年左右，统计学才开始“从他所谓的二战后‘过度数学化’中‘恢复’”。51伴随着这种实践上的变化，他描述了一种根本性的对比：一种是“数据建模文化”，估计“98% 的统计学家”都在使用这种文化；另一种是“算法建模文化”，只有“2% 的统计学家”但在“其他领域也有很多人”在使用。在主导学术统计学的数据建模文化中，模型验证是通过“使用拟合优度检验是或否”和“使用算法建模文化”来实现的。残差检查”。相比之下，算法文化则专注于“预测准确性”。52将自己限制在当代统计学的有限模型范围内意味着放弃大量数据，要求比通常可能的更多的确定原因知识，并限制创造解决当代问题所需的新工具。算法文化可以提供的东西太多了，即使这意味着放松统计学的传统要求。
Statistics was born from making sense of data about diverse populations and systems analyzing data, and yet in the eyes of practitioners like Breiman, the discipline had gone far astray; only then, around 2000, was statistics beginning “to ‘recover’ from what he called its ‘overmathematiza-tion’ in the post–World War II years.”51Attendant upon this change in practice, he described a radical contrast between a “data modeling culture” used by an estimated “98% of all statisticians” and an “algorithmic modeling culture,” used by “2% of statisticians” but “many in other fields.” In the data modeling culture dominating academic statistics, model validation comes through “Yes-no-using goodness-of-fit tests and residual examination.” In contrast, the algorithmic culture focused on “predictive accuracy.”52 Restricting oneself to the limited range of models of contemporary statistics was to abandon vast arrays of data, to demand more certain knowledge of causes than often possible, and to limit the creation of new tools needed to solve contemporary problems. Algorithmic culture had too much to offer, even if it meant loosening the traditional demands of statistics.
布雷曼并不是唯一一个呼吁数理统计学重新将焦点转向现实世界数据的人，现在有了数字计算机的帮助。20 世纪 70 年代末，其他统计学家呼吁他们的领域更充分地利用数字计算机提供的可能性。尽管计算能力不断增长，但布雷曼、布拉德利·埃夫隆和威廉·克利夫兰等从业者认为，学术统计学家未能正视庞大的现实世界数据集，未能将计算更集中地融入他们对该领域的理解中。1993 年，贝尔实验室的约翰·钱伯斯呼吁建立一种从数据中学习的“更大的统计学”，他担心统计学过于孤立的数学驱动力“限制了统计学的影响力和该领域为社会带来的好处”。53数据的爆炸式增长为统计学提供了机会，使其能够在确保严谨性和启发新方法方面发挥重要作用，但该领域未能利用这种可能性。两位主要的计算导向统计学家 Walter Stuetzle 和 David Madigan 呼吁彻底颠覆统计学研究生教育，注重不同的学科特性。
Breiman was not alone in calling for mathematical statistics to return its focus to real-world data, now with the help of digital computers. In the late 1970s, other statisticians called for their field to more fully embrace the possibilities the digital computer afforded. Despite the growth of computational power, practitioners such as Breiman, Bradley Efron, and William Cleveland argued that academic statisticians failed to face up to large real-world data sets and to integrate computing more centrally within their understanding of the field. In his 1993 call for a “greater statistics” that would learn from data, John Chambers of Bell Labs worried that the overly insular mathematical drive of statistics was “limiting both the influence of statistics and the benefits the field had provided to society.”53 The explosion of data had created the opportunity for statistics to serve an essential function in ensuring rigor and inspiring new methods, but the field was failing to leverage that possibility. Two major computationally oriented statisticians, Walter Stuetzle and David Madigan, called for a dramatic upending of graduate education in statistics, focused on different disciplinary identity.
统计学主要侧重于从有限的数据中榨取最大信息量。这一模式的重要性正在迅速下降，统计学教育也发现自己与现实脱节。54
Statistics has primarily focused on squeezing the maximum amount of information out of limited data. This paradigm is rapidly diminishing in importance and statistics education finds itself out of step with reality.54
如果愿意的话，统计学家可以为机器学习和数据挖掘做出很多贡献。
Statisticians had much to offer to machine learning and data mining, if they would let themselves.
统计学家 Bin Yu 在“让我们拥有数据科学”的演讲中提到，统计部门注意到了数据科学的兴起。她认识到大学里教授的统计学世界与数据科学世界之间存在着巨大的鸿沟，她呼吁做的不仅仅是重塑品牌。“数据科学代表着大数据时代计算和统计思维的必然（重新）融合。我们（统计学家）必须拥有数据科学，因为领域问题并不区分计算和统计，反之亦然，数据科学是处理现代数据问题的新术语。”
Statistics departments took note of the rise of data science, as suggested by the statistician Bin Yu in her talk “Let Us Own Data Science.” Recognizing the gulf between the world of statistics as taught in universities and the world of data science, she called for far more than an exercise in rebranding. “Data Science represents an inevitable (re)-merging of computational and statistical thinking in the big data era. We [statisticians] have to own data science, because domain problems don’t differentiate computation from statistics or vice versa, and data science is the new accepted term to deal with a modern data problem in its entirety.”
“数据科学家的崛起”
“The Rise of the Data Scientist”
杰夫·哈默巴赫 (Jeff Hammerbacher) 在 2009 年写下了他对数据科学家的描述，其中融合了克利夫兰 2001 年提案的思维方式、20 世纪 90 年代数据挖掘的商业规模以及 21 世纪初迅速兴起的“大数据”民主化工具集。思维方式和工具集都源自行业经验，Facebook 在其早期发展阶段的经验、贝尔实验室团队自数字计算诞生之初就致力于通过数据理解世界（并提高利润）的经验，以及在日常业务工作中扩展数据分析的众多努力。
When Jeff Hammerbacher wrote his description of data scientists in 2009, it combined the mindset of Cleveland’s proposal of 2001 and the commercial scale of 1990s data mining with a rapidly emerging toolset for democratizing “big data” in the early 2000s. Mindset and toolset alike were informed by industrial lessons learned, at Facebook in its early days of growth, teams at Bell Labs working since the dawn of digital computation to make sense of the world (and advance the bottom line) through data, and the many efforts to scale data analysis in everyday business work.
2000 年初，云计算的成本也大幅下降，而云计算本身也得益于互联网的信息基础设施，它允许数据从世界各地的计算机来回流到数千英里之外的计算中心。这与 20 世纪 90 年代的数据挖掘时代相呼应，这鼓励公司将他们的网络日志、商业交易流和客户记录进入数据存储，希望可以“挖掘”其中的利润模式。例如，数十年来在适用于健康或金融数据的行业监管中发展起来的消费者数据保护措施，对许多在线公司的做法几乎没有任何权威性，即使在个人数据被采集和分析时也是如此。云计算给公司带来的另一个好处是隐形或“幽灵劳动力”，正如刚才讨论的那样，实际的工人可能在地球上的任何地方。
The beginning of the millennium also gave rise to a dramatic drop in the cost of computation via cloud-hosted computing, itself facilitated by the information infrastructure of the internet, which allowed data to flow from a computer anywhere in the world back and forth to compute centers thousands of miles away. Echoing the data mining moment of the 1990s, this encouraged companies to turn their web logs, streams of commercial transactions, and customer records into stores of data in the hopes that they could be “mined” for profitable pattern discovery. Protections for consumer data, developed over decades in sectoral regulation applicable to health or financial data, for example, had little authority over the practices of many online companies, even when personal data was being ingested and analyzed. An additional benefit to companies from the cloud was the invisible or “ghost labor,” just discussed, where the actual worker could be anywhere on the planet.
面向行业的出版物，如《哈佛商业评论》和 O'Reilly Media 开始大肆赞扬应用于这些大量交易数据的机器学习方法，并承诺为那些精炼和加工新石油的人带来财富和颠覆。数据挖掘、大数据和预测分析的早期进步（和营销）促进了行业对数据科学的热情采用。很快，各种服务提供商和初创公司应运而生，向新矿工出售数字镐，数据科学的福音充斥着他们的营销材料，转动数据繁荣的飞轮，鼓励新老公司重新考虑他们的数据战略和人员配置。55
Industry-facing publications such as Harvard Business Review and O’Reilly Media began to sing the praises of machine learning methods applied to these large troves of transactional data, promising riches and disruption to those who would refine and process the new oil. The enthusiastic adoption of data science in industry was facilitated by the earlier advances in (and marketing of) data mining, big data, and predictive analytics. In short order, a variety of service providers and start-up companies were born to sell digital pickaxes to the new miners, and the gospel of data science filled their marketing materials, turning the flywheel of data exuberance and encouraging companies old and new to reconsider their data strategy and staffing.55
由于硅谷的初创企业，数据科学家成为一种明确的工作描述，如今，尽管有时不情愿，它已融入研究和高等教育的结构中。在一些大学，数据科学是一个新的机构；在其他大学，它作为现有部门的更名而蓬勃发展。长期以来，根据 Tukey、Cleveland、Breiman 等人的宣言，一些统计部门开始将自己更名为统计和数据科学部门，例如耶鲁大学和卡内基梅隆大学，都是在 2017 年更名。
The data scientist became a clear job description thanks to start-ups of Silicon Valley, and has now become woven, at times begrudgingly, into the fabric of research and higher education. In some universities, data science is a new institute; in others, it thrives as a renaming of existing departments. Long cold to the manifestos of Tukey, Cleveland, Breiman, and others, some statistics departments began to rename themselves as departments of statistics and data science, as at, for example, Yale and Carnegie Mellon, both in 2017.
根据对数据科学实践团队的民族志观察，Gina Neff 及其合著者认为，“理解数据是一个集体过程。” 56在某些方面，这满足了钱伯斯 1993 年梦想“更大”统计学的要求，尽管有可能成为一个关于一切的领域，正如詹妮弗·布莱恩和哈德利·威克姆在《数据科学：一个三环马戏团还是一个大帐篷？》57中警告的那样，数据科学家这一职位描述也证明了自己是一个不断变化的目标。Reddit 上“数据科学”子版块的一篇帖子问道：“Facebook 的数据科学家真的是数据分析师吗？” 58随着数据实践变得越来越专业化，这种变化也带来了职位名称的激增。今天，不仅有数据分析师和数据科学家，还有数据工程师、分析工程师，以及反映数据日益增长的政策和道德影响的“数据治理”专业职能。
Based on their ethnographic observations of practicing data science teams, Gina Neff and coauthors argue, “making sense of data is a collective process.”56 In some ways this has met Chambers’s 1993 dream of “greater” statistics, though at the risk of being a field about and of everything, as Jennifer Bryan and Hadley Wickham warned in their “Data Science: A Three Ring Circus or a Big Tent?”57 As with the terms “artificial intelligence” and “machine learning,” the job description “data scientist” proved itself to be a moving target. A Reddit post in the “data science” subreddit asks: “Are Data Scientists at Facebook really Data Analysts?”58 Along with this drift came a proliferation of job titles as the practice of data became ever more specialized. Today one has not only data analysts and data scientists, but also data engineers, analyst engineers, and, reflecting the growing policy and ethical implications of data, the professional function of “data governance.”
数据科学不仅包括技术工具的民主化——如今高级统计软件和强大的计算能力随处可见——还包括技能的民主化。正如 20 世纪 50 年代以来许多社会科学家不加批判地使用 p 值一样，许多学科的研究人员也开始使用各种数据科学工具，但并不总是小心谨慎。然而，这些技术的易用性使得许多工作远非反思性、批判性或变革性。
Data science has come to encompass a democratization not only of technological tools—high-level statistical software and powerful computing are now easily available—but equally of skills. Just as too many social scientists uncritically used p values from the 1950s forward, researchers in many disciplines have begun using the full range of data science tools, but not always with care. The ease of use of these technologies allows, however, much work that is far from reflexive or critical or transformative.
COVID-19 的挑战促使世界各地的研究人员尝试应用机器学习来预测疫情的进程，但结果好坏参半。卡内基梅隆大学统计与数据科学系的 Ryan Tibshirani 在对该领域的预测能力（包括他自己的工作）进行了批判性评估后得出结论：“作为一个社区，我们错过了每一次激增（意思是没有预料到它们）。” 59更成问题的是，2017 年左右，长期被嘲笑的通过分析面部特征来确定人类特征的“科学”——相貌学——以机器学习研究的形式回归。60统计学一直与人类差异分析有着密切而又常常令人困扰的关系。尽管这些伪科学遭到了严厉的批评，但自 Quetelet 和 Galton 以来，研究人员一直试图使用统计方法对人进行分类，并且他们长期以来一直在寻找技术手段来区分先天和后天——以发现真正的天才——以及真正的罪犯。引人注目的头条新闻经常刊登与伪科学相貌学近乎的机器学习预印本。Luke Stark 和 Jevan Hutson 认为，“人工智能和机器学习现在可以预测你是否会犯罪、是否是同性恋、是否会成为一名好员工、你是政治自由派还是保守派，以及你是否是精神病患者，所有这些都基于你的面部、身体、步态和语调等外部特征。” 61这些作品不仅滥用机器学习，还为旧式的科学种族主义披上了新的客观外衣。
The challenges of COVID-19 encouraged researchers worldwide to attempt to apply machine learning to predicting the course of the pandemic, with mixed results. In a critical assessment of the field’s predictive power, including his own work, Ryan Tibshirani of Carnegie Mellon’s Department of Statistics and Data Science concluded, “We as a community missed every surge (meaning, didn’t anticipate them).”59 In a more problematic vein, around 2017 the long-derided “science” of determining human character through the analysis of the face—physiognomy—returned in the form of machine learning studies.60 Statistics has always had a close and often troubled relationship to the analysis of human difference. Despite devastating criticism of these pseudosciences, researchers since Quetelet and Galton have sought to classify people using statistical means, and they have long sought technical means to distinguish nature from nurture—to find the true geniuses—and the true criminals. Attention-grabbing headlines regularly feature machine learning preprints bordering on pseudoscientific physiognomy. Luke Stark and Jevan Hutson argue, “artificial intelligence and machine learning can now purportedly predict whether you’ll commit a crime, whether you’re gay, whether you’ll be a good employee, whether you’re a political liberal or conservative, and whether you’re a psychopath, all based on external features like your face, body, gait, and tone of voice.”61 Not just making poor use of machine learning, these works give old-style scientific racism a new objective sheen.
重点不是谴责数据科学工具，而是更恰当地使用它们，并意识到它们的局限性。当马哈拉诺比斯改进了皮尔逊的工具来研究印度的种姓制度时，他对人们可能得出的结论采取了更为批判的态度。2010 年代初，我们俩都有幸在哥伦比亚大学教授首届数据新闻学课程，我们试图向来自世界各地的优秀年轻记者传授批判性使用数据科学技术的方法，通过仔细的数据分析和收集、算法分析和可视化来检查政府和公司；我们对这些工具如何实现和增强批判性调查工作的方式持乐观态度。学者凯瑟琳·迪格纳齐奥和劳伦·克莱因在《数据女权主义》一书中阐述了研究人员如何批判性地使用数据科学工具来发挥解放潜力，而不是再次以新的科学外衣重新讨论糟糕的旧伪科学。62
The point is not to condemn data science tools—it’s to use them more appropriately, and with an appreciation for their limits. When Mahalanobis improved Pearson’s tools in looking at caste in India, he adopted a far more critical approach to the conclusions one could reach. In the early 2010s, both of us had the privilege of teaching an inaugural program in data journalism at Columbia where we sought to teach astonishing young journalists from around the world the critical use of data science technology, to check governments and corporations alike, using careful data analysis and collection, algorithmic analysis, and visualization; we have optimism for the way these tools enable and empower critical investigative work. In their Data Feminism, the scholars Catherine D’Ignazio and Lauren Klein illustrate how researchers can make critical use of data science tools to have liberatory potential, rather than rehashing bad old pseudosciences in new scientific garb yet again.62
随着数据科学的范围不断扩大，认识到数据不仅可以用于下棋和围棋，或区分狗和猫的照片，还可以用于解决危害和正义受到威胁的人类问题，从而发挥强大的权力。换句话说，在数据科学的旗帜下，“通过数据理解世界”的扩展部分在于人们越来越认识到数据的伦理、政治和社会影响。
As the scope of data science has expanded, so has the realization that data can be a powerful force when applied not only to playing games of chess and go, or distinguishing photographs of dogs and cats, but when applied to human problems when harms and justice are at risk. Said otherwise, part of the expansion of “making sense of the world through data” under the banner of data science has been the increasing recognition of data’s ethical, political, and social impacts.
没有专业知识的道德
Ethics without Expertise
1997 年 11 月 10 日，当时还是研究生的谢尔盖·布林（后来成为谷歌联合创始人）在一封邮件中写道：“作为数据挖掘领域的科学家，我们必须定期从技术中抽身，思考使用它的道德问题。”他举了几个例子：
“As scientists in the field of data mining,” Sergey Brin, then a graduate student, later to co-found Google, wrote a listserv on November 10, 1997, “it is important for us to periodically take a step back from the technology and consider the ethics of using it.” He offered a few examples:
汽车保险公司分析事故数据，并根据年龄、性别、车辆类型等设定个人保险费率。如果法律允许，他们还会使用种族、宗教、残疾以及他们发现的与事故率相关的任何其他属性。健康保险公司也使用类似的数据。所有这些都可以看作是数据挖掘的结果，它们对人们的生活有着重大影响。63
auto insurance companies analyse accident data and set insurance rates of individuals according to age, gender, vehicle type, . . . If they were allowed to by law, they would also use race, religion, handicap, and any other attributes they find are related to accident rate. Health insurance companies also use similar data. . . . All of these can be seen as results of data mining and they have a significant affect [sic] on people’s lives.63
他要求同事们“请提出您的意见和任何相关的例子或研究。”数据挖掘以及后来的数据科学都是高度跨学科的；它们对需要哪些专业知识有限制。虽然斯坦福会议通常是跨学科的，但似乎没有受过道德训练的人参加，就像建筑工程师或生物学家参加专门讨论他们学科的会议一样。人们只能想知道讨论的结论是否是“不作恶”。无论考虑得多么周全，意图多么良好，道德往往都无法很好地扩展。斯坦福数据挖掘文化中的人们非常了解如何扩展算法；他们知道如何制定行业；他们知道如何推动学术研究走向实用目的。至于扩展道德，公平地说，这方面的讨论较少。
He asked his colleagues to “Please bring your opinions and any relevant examples or studies.” Data mining, and later data science, were highly interdisciplinary; they had limits to whose expertise was called upon. While the Stanford meetings were often interdisciplinary, it doesn’t appear anyone trained in ethics attended, any more than construction engineers or biologists attended sessions dedicated to their disciplines. One can only wonder if the conclusion of the discussion was “Don’t be evil.” Ethics, no matter how well considered and well intentioned, tends not to scale well. People in the data mining culture at Stanford knew well how to scale algorithms; they knew how to draw up industry; they knew how to prompt academic research toward practical ends. As for scaling ethics, that was, fair to say, less well covered.
* Grus，《数据科学之路》。他引用了罗伯特·海因莱因的话：“人类应该能够换尿布、策划入侵、杀猪、驾驶轮船、设计建筑、写十四行诗、核算账目、砌墙、接骨、安慰垂死之人、接受命令、发布命令、合作、单独行动、解方程、分析新问题、施肥、编写计算机程序、烹饪美味佳肴、高效战斗、英勇牺牲。专业化是昆虫的专利。”罗伯特·A·海因莱因，《爱情的时间足够：拉撒路·朗的生活；一部小说》（纽约：普特南出版社，1973 年）。
* Grus, “The Road to Data Science.” He references Robert Heinlein: “A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently, die gallantly. Specialization is for insects.” Robert A. Heinlein, Time Enough for Love: The Lives of Lazarus Long; a Novel (New York: Putnam, 1973).
在我们的文化传统中普遍接受的原则中，有三项基本原则与涉及人类受试者的研究伦理特别相关：尊重他人、仁慈和正义的原则。
Three basic principles, among those generally accepted in our cultural tradition, are particularly relevant to the ethics of research involving human subjects: the principles of respect of persons, beneficence and justice.
–《贝尔蒙特报告》，1978 年
–The Belmont Report, 1978
如今，我在人工智能文献中经常看到的是“道德”。我想扼杀道德。
What I always see in the AI literature these days is “ethics.” I want to strangle ethics.
–Philip G. Alston，纽约大学 John Norton Pomeroy 法学教授，AI Now 2018 研讨会1
–Philip G. Alston, John Norton Pomeroy Professor of Law, NYU, AI Now 2018 Symposium1
2020年初，谷歌组建了一支人工智能伦理研究团队，由该领域两位杰出的早期职业学者领导。玛格丽特·米切尔博士和蒂姆尼特·格布鲁博士凭借重要的学术和流行出版物，以阐明人工智能的潜在和实际危害以及提出减轻这些危害的建设性方法而闻名。除了其他开创性的工作外，格布鲁早些时候与乔伊·布奥拉姆维尼博士一起表明，几种常见的“商业性别分类系统”在对不同人口群体进行分类的准确性方面存在很大差异，尤其是“深色皮肤的女性……最容易被误分类的群体”。2米切尔因项目而闻名关于“消除偏见”机器学习的研究，以及与 Gebru 的合作，包括他们在“模型报告的模型卡”方面的工作……这是朝着机器学习和相关人工智能负责任的民主化迈出的一步。” 3到 2020 年夏天，该公司已准备好提供谷歌的 AI 伦理方法作为一项服务：“谷歌提供帮助他人解决棘手的 AI 伦理问题”，据《连线》杂志的一篇标题报道，“在经历了艰难的道德教训后，这家科技巨头将提供诸如发现种族偏见或围绕 AI 项目制定指导方针等服务。” 4
In early 2020, Google formed an AI ethics research team, led by two prominent early career scholars in the field. With major academic and popular publications, Dr. Margaret Mitchell and Dr. Timnit Gebru were known for illustrating the potential and real harms of artificial intelligence and suggesting constructive ways to mitigate these harms. Among other groundbreaking work, Gebru had earlier shown, along with Dr. Joy Buolamwini, that several common “commercial gender classification systems” exhibit “substantial disparities in the accuracy of classifying” different demographic groups, particularly “darker-skinned females . . . the most misclassified group.”2 Mitchell was well known for projects on “debiasing” machine learning, as well as collaborations with Gebru including their work on “Model Cards for Model Reporting . . . as a step towards the responsible democratization of machine learning and related artificial intelligence.”3 By summer 2020, the company was ready to offer Google’s approach to AI ethics as a service: “Google Offers to Help Others With the Tricky Ethics of AI,” according to a headline in Wired, “After learning its own ethics lessons the hard way, the tech giant will offer services like spotting racial bias or developing guidelines around AI projects.”4
谷歌已经成功解决了许多以前无法克服的数据问题——搜索、计算机视觉，甚至机器翻译。它很快就会在这个棘手的问题上取得进展吗？随着 2020 年接近尾声，一个开明的人工智能道德团队和谐地融入公司决策框架的愿景已经破灭。同年 11 月，格布鲁宣布她已被谷歌解雇；2021 年初，米切尔也发表了类似的声明。谷歌声称，格布鲁因一项研究出版物质量纠纷而辞职；她反驳说，她被解雇的原因是要求谷歌承认大型语言模型（其核心技术之一）可能带来的道德危害。5这些研究人员和他们的前雇主发表的一系列公开声明暴露了人工智能研究人员（即使是在同一家公司）对公司将道德融入人工智能研究的期望与一家公司产品开发的现实之间存在着多么巨大的差距，而这家公司的盈利模式正是建立在大量使用人类数据的基础上。
Google had succeeded in the face of many previously insurmountable data problems—search, computer vision, even machine translation. Might it soon make progress on this thorny topic? As 2020 came to a close, the vision of an enlightened AI ethical team, harmoniously integrated into the decision-making framework of the corporation, had collapsed. In November of that year, Gebru announced that she had been fired by Google; early in 2021, Mitchell made a similar announcement. Google claimed that Gebru had resigned over a dispute concerning the quality of a research publication; she countered that she was fired for demanding that Google admit to the potential ethical harms of large language models, one of its core technologies.5 A flurry of public pronouncements by these researchers and their former employer exposed just how vast the gulf between ways AI researchers, even in the same company, might expect a company to integrate ethics in AI research and the realities of product development in a corporation whose profit model rested precisely on the massive use of data on people.
“IRB 的演变”
“Evolving the IRB”
这不是平台公司第一次因道德失误而受到公众的审查。一个有启发性的案例研究2014 年，Facebook 研究人员发表了“情绪感染”研究论文，这一天到来了。6媒体对此反应强烈，标题包括“Facebook 故意让人伤心。这应该是最后一根稻草” 7和“用户对 Facebook 情绪操纵研究感到愤怒”。8更让Facebook 领导层担忧的是电子隐私信息中心向联邦贸易委员会 (FTC) 提出的投诉，以及参议员马克·沃纳正式要求 FTC 调查这项研究。突然之间，快速行动和打破常规导致了限制利润的监管的可能性。
This was not the first time a platform company had endured public scrutiny over ethical lapses. An instructive case study arrived in 2014 with the publication of the “emotional contagion” research paper by Facebook researchers.6 Negative reaction in the press was damning, with headlines such as “Facebook Deliberately Made People Sad. This Ought to Be the Final Straw”7 and “Users Angered at Facebook Emotion-Manipulation Study.”8 Of more material concern to Face-book leadership were a Federal Trade Commission (FTC) complaint filed by the Electronic Privacy Information Center and a formal request by Senator Mark Warner that the FTC investigate the research. Suddenly, moving fast and breaking things had led to the possibility of profit-curtailing regulation.
Facebook 采取了先发制人的自我监管举措：明确将研究伦理引入 Facebook，通过“发展”学术界基于原则的机构审查委员会 (IRB) 流程以适应企业环境。由于大型信息平台公司中的许多研究人员都接受过学术培训，IRB 的应用伦理概念——围绕由裁决机构解释的综合原则——影响了技术界许多人的思维。正如社会学教授、普林斯顿大学 IRB 前成员马特·萨尔加尼克 (Matt Salganik) 所写，“基于原则的方法足够普遍，无论你在哪里工作（例如大学、政府、非政府组织或公司），它都会有所帮助。” 9
Facebook responded with a preemptive, self-regulatory move: explicitly bringing research ethics to Facebook by “evolving” academia’s principles-based institutional review board (IRB) process for the corporate environment. Since many researchers within large information platform companies come from academic training, the IRB conception of applied ethics—around comprehensive principles which are interpreted by an adjudicating body—has colored the thinking of many in the technology community. As Matt Salganik, professor of sociology and former member of Princeton’s IRB, writes, “the principles-based approach is sufficiently general that it will be helpful no matter where you work (e.g., university, government, NGO, or company).”9
为了回应公众对情绪感染研究的争论，Facebook 的 Molly Jackman 和 Lauri Kanerva 于 2016 年发表了《IRB 的发展：为行业研究建立强有力的审查》，记录了一种扩展机构审查委员会应用伦理的方法。基于 Kanerva 在斯坦福大学领导非医学 IRB 的十年经验，该文章为 Facebook 制定了组织审查委员会流程。可以肯定的是，创建 IRB 并不能治愈所有疾病，也不能阻止所有道德问题争论。事实上，在情绪感染研究中，大学 IRB 认为该研究无需审查，因为它不涉及“人类受试者”，至少没有以医学和社会科学研究习惯的研究审查委员会所理解的方式可见的受试者。然而，这一决定随后引发了如此巨大的争议，以至于世界最重要的科学杂志之一《国家科学院院刊》的编辑们罕见地发表了关于该论文的“道德关切声明” 。10
In response to public debate about the emotional contagion study, Facebook’s Molly Jackman and Lauri Kanerva published in 2016 “Evolving the IRB: Building Robust Review for Industry Research,” documenting an approach expanding the applied ethics of institutional review boards. Building on Kanerva’s ten years of experience leading the nonmedical IRB at Stanford, the article set out an organizational review board process for Facebook. To be sure, creating an IRB does not cure all ills and prevent all ethical debates. In the emotional contagion study, in fact, a university IRB had deemed the research not necessary to review as it involved no “human subjects,” at least none visible in the way understood by research review boards accustomed to medical and social scientific research. So great, however, was the ensuing relitigation of this decision that the editors of PNAS, one of the foremost scientific journals in the world, issued a rare “statement of ethical concern” about the paper.10
尽管人们曾多次尝试定义伦理原则，但没有一项尝试能像人类受试者研究方法那样产生如此大的影响。IRB 正是在这一领域首次被提出，并且是在对科学丑闻的反应中发展起来的。为了理解计算社会科学家用来构建应用伦理学的背景，并与主流科技公司最近提供的各种原则、立场和产品进行对比，有必要重新回顾 20 世纪 70 年代《贝尔蒙特报告》中的应用伦理学起源。该报告是定义应用伦理学的基础文件，以及由此产生的伦理审查流程，即机构审查委员会的设立（以及受 IRB 启发的组织审查委员会，如 Facebook 的设立）。这种伦理制度化为最近关于如何定义数据赋权算法伦理的讨论提供了一个具有影响力的背景。
While there have been many attempts to define principles for ethics, none has had near the impact of the human subjects research approach, the domain in which IRBs were first proposed, and itself developed in reaction to scientific scandal. To understand the context computational social scientists have used to frame applied ethics, and to contrast with the diverse set of principles, postures, and products recently offered by dominant technology companies, it’s useful to revisit its origin in the 1970s in the form of the Belmont Report. This report serves as a foundational document for defining applied ethics, along with the resulting process for ethical review, namely, the creation of institutional review boards (and thus the IRB-inspired organizational review boards such as Facebook’s). This institutionalization of ethics has been an influential backdrop for more recent discussions of how to define the ethics of data-empowered algorithms.
从塔斯基吉到贝尔蒙特
From Tuskegee to Belmont
通往贝尔蒙特的道路是由美国公共卫生服务部的医生和科学家的研究抱负铺就的；它最终登上了《纽约时报》的头版，被揭露为一种道德上的失败，这种失败是如此的种族主义和科学上的缺陷，以至于一个跨学科团队花了数年时间才制定出联邦立法回应，旨在确保纳税人的钱不会再用于资助这样的灾难。11
The road to Belmont was paved with the research aspirations of the doctors and scientists of the US Public Health Service; it would end on the front page of The New York Times, exposed as an ethical failure so racist and scientifically flawed that it would launch years of work by an interdisciplinary team to craft federal legislative response aimed at ensuring that taxpayer dollars would never again fund such a catastrophe.11
1973 年 7 月 26 日，《美国公共卫生服务局在塔斯基吉开展的梅毒研究》登上了《纽约时报》的头版，标题为《美国研究中的梅毒受害者 40 年未接受治疗》。美国公众了解到，数十年来，纳税人的钱一直支持着一项系统性地拒绝阿拉巴马州塔斯基吉市非裔美国男性接受梅毒治疗的研究。这项实验在科学上毫无用处，而且带有强烈的种族主义色彩。塔斯基吉的结束恰逢人们对美国政府极度不信任的时期。
On July 26, 1973, the “U.S. Public Health Service Syphilis Study at Tuskegee” made the front page of The New York Times, with the headline “Syphilis Victims in U.S. Study Went Untreated for 40 Years.” The American public came to know that taxpayer funding had, for decades, supported a study that systematically denied African American men in Tuskegee, Alabama, treatment for syphilis. The experiment was scientifically useless as well as deeply racist. The end of Tuskegee coincided with a time of deep distrust in the US government.
1936 年 5 月，美国医学协会会议以黑人男性梅毒未治疗为主题。用美国公共卫生服务部医疗主管的话来说，塔斯基吉地区一群未治疗的人群“似乎提供了一个难得的机会，可以研究未治疗的梅毒患者从发病开始到死亡的全过程”。12结果很明显：治疗产生了显著的积极效果。抓住这些实验“机会”意味着数十年的痛苦。未治疗的男性只能忍受这种状态——接受安慰剂治疗，几十年内无法接受治疗，甚至在二战期间被推迟服兵役，以阻止他们接受治疗。13在医学界广为人知的“塔斯基吉研究”中，一种优先考虑潜在科学知识的不可抗拒的科学逻辑与正义的伦理考虑和对个人知情自主权的尊重相冲突。发起并继续实验的决定也反映了 20 世纪结构性种族主义和优生思想的长期影响。该项目持续进行，定期发布其结果，直到 1972 年，一名意识到纳粹和日本战时实验的告密者将此事推向公众视野。
The American Medical Association meeting of May 1936 featured a lecture on untreated syphilis in Black American men. A population of untreated people from the Tuskegee area, in the words of the medical director of the US Public Health Service, “seemed to offer an unusual opportunity to study the untreated syphilitic patient from the beginning of the disease to the death of the infected person.”12 The results were clear: treatment has dramatically positive effects. Taking these experimental “opportunities” meant decades of suffering. The untreated men were kept that way—given placebo treatments, prevented from getting treatment for several more decades, even given draft deferments to keep them from getting treatment during World War II.13 In the “Tuskegee study,” as this effort became known in the medical community, an inexorable scientific logic, prioritizing potential scientific knowledge, clashed with ethical considerations of justice and respect for individual’s informed autonomy. The decisions to initiate and continue the experiment reflected as well the long-term grip of structural racism and eugenic thought in the twentieth century. The project continued, publishing its results regularly, until 1972, when a whistleblower, conscious of Nazi and Japanese wartime experimentation, pushed the story into the public eye.
在该事件引起轰动后，美国国会成立了一个委员会，“以确定在涉及人类受试者的生物医学和行为研究中应遵循的基本道德原则”。委员会成员来自不同领域，包括研究人员、律师、哲学家和一位前天主教神父。14他们负责制定研究道德框架，并设计一个流程，以确保该框架能够指导和约束研究人员的行为。最终的报告确立了一种结合学术哲学、社会规范和研究过程现实的道德方法。尽管他们当时的动机问题（例如对儿童的研究、胎儿研究和对囚犯的研究）似乎与数据赋能的算法决策系统的问题不同，但该委员会旨在提供一个更普遍的研究框架。该报告于 1979 年 4 月 18 日登上《联邦公报》，距离塔斯基吉大学首次被新闻曝光已过去六年。
In the wake of its explosive recognition, the US Congress set up a commission “to identify the basic ethical principles that should underlie the conduct of biomedical and behavioral research involving human subjects.” The diverse group of commissioners included researchers, lawyers, philosophers, and a former Catholic priest.14 They were charged with devising an ethical framework for research as well as with designing a process to ensure that this framework would guide and constrain the behavior of researchers. The resulting report established an approach to ethics that combined academic philosophy, social norms, and the realities of the research process. While their motivating problems of the day such as research on children, fetal research, and research on the incarcerated may seem different from those of data-empowered algorithmic decision systems, the commission aimed to provide a framework useful in research more generally. The report entered the Federal Register on April 18, 1979, almost six years after the first journalistic exposés on Tuskegee.
报告坚持认为，涉及人的研究不能再仅仅以其对整个社会的长期利益为依据。研究方案必须仔细权衡研究对每个参与者的影响：“影响直接研究对象的风险和利益通常会具有特殊意义。”委员会警告不要利用受压迫和无权无势的群体：
The report insisted that research studies involving people could no longer simply be justified by their claimed long-term benefits to society as a whole. Research protocols must carefully weigh the impact of the study on each person involved in it: “the risks and benefits affecting the immediate research subject will normally carry special weight.” And the commission warned about taking advantage of oppressed and disempowered groups:
某些群体，如少数族裔、经济弱势群体、重病患者和被收容在机构的人，可能会不断被寻求成为研究对象，因为他们在进行研究的环境中随时可用。鉴于他们的依赖地位和他们经常受到损害的自由同意能力，他们应该受到保护避免仅仅为了行政上的方便而参与研究，或因为他们的疾病或社会经济状况而容易被操纵。15
Certain groups, such as racial minorities, the economically disadvantaged, the very sick, and the institutionalized may continually be sought as research subjects, owing to their ready availability in settings where research is conducted. Given their dependent status and their frequently compromised capacity for free consent, they should be protected against the danger of being involved in research solely for administrative convenience, or because they are easy to manipulate as a result of their illness or socioeconomic condition.15
该报告限制了对人类的研究，并建立了一套强有力的机构来执行这些限制，尽管这些限制并不完善。要使伦理道德得以坚持，意味着政府首先必须批准一些伦理道德，然后设计一个由法律和强有力的官僚机构执行的流程，以指导和约束伦理研究，并制裁滥用和滥用行为。
The report led to limits to research on human beings and a robust set of institutions that implement those limits, however imperfectly. Making ethics stick meant the government first had to sanction some account of ethics, and then design a process, enforced by law and robust bureaucracies, that would guide and constrain ethical research and sanction misuse and abuse.
由此产生的伦理实验框架被记录在《贝尔蒙特报告》中。16委员们定义的伦理学强调了手段与目的之间的矛盾（或者，作为哲学框架，义务论与结果论之间的矛盾） ，并坚持正义，包括公平分配社区之间的利益和损害。 《贝尔蒙特报告》没有设定一套具体的规则或单一的准则，而是将伦理学作为这些矛盾的协商解决方案，并以三项原则作为共同的认知支撑——各方即使在具体应用上存在分歧时也能达成共识：
The resulting framing of ethical experimentation was captured in the Belmont Report.16 Ethics, as defined by the commissioners, enshrined the tension between means against ends (or, as philosophical frameworks, deontology against consequentialism) and insisted on justice, including the fair allocation of benefits and harms across communities. Rather than setting a specific set of rules or a single maxim, the Belmont Report sets out ethics as a negotiated resolution of these tensions, with three principles as the shared epistemic backstop—the consensus on which all parties can agree, even when disagreeing about specific applications:
1.尊重人格：尊重个人的自主权；
1. Respect for personhood: the idea that individuals’ autonomy should be respected;
2.仁慈：最小化个人受到伤害的风险，最大化公共利益；
2. Beneficence: minimize risk of harm to individuals, maximize public benefit;
3.公正：风险和利益的公平分配。
3. Justice: fair distribution of risk and benefits.
在流行文化中，伦理学通常被认为是一种哲学论证，或者可能是一份小清单，甚至可能是一条准则。然而，贝尔蒙特团队的方法却是“原则主义”。原则主义的理念是定义一小组具有足够普遍性的原则，使它们不仅适用于当前问题，而且可能适用于未来问题。在《贝尔蒙特报告》中，作者明确表示，他们的目标是让这些原则“全面”，这意味着他们预计这些原则将适用于未来人类受试者研究中所有应用伦理问题。但每个案例都不同。任何一套原则如何适用？
In popular culture, ethics is often conceived of as a philosophical argument or perhaps a small checklist of items, or perhaps even a single maxim. However, the approach of the Belmont group was instead one of “principlism.” The idea of principlism is to define a small set of principles with enough generality that they will be applicable not only to the present concerns but likely to future concerns. In the Belmont Report itself, the authors explicitly state that they aim their principles to be “comprehensive,” meaning that they anticipate their utility for all future applied ethics problems in human subjects research. But every case is different. How can any set of principles apply?
与美国宪法等管理性文件一样，记录下来的一系列原则的价值在于社区，社区必须努力将这些原则解释为更符合具体情况的标准，并最终制定针对具体案例的明确规则。与宪法一样，贝尔蒙特报告本身也是一种通用指南，组织或社区中的每个人都可以同意其合法性。但是，该文件的效力和实用性受到社区存在的限制，社区需要努力将这些原则提炼为标准、规则，从而付诸实践。
As with a governing document such as the United States Constitution, the value of the documented set of principles is in the community that must strive to interpret these principles as more context-specific standards, and eventually to create unambiguous rules specific to individual cases. Like the Constitution, the Belmont Report itself functions as a guide so general that everyone in an organization or a community can agree on its legitimacy. But the power and utility of the document is limited by the existence of a community which does the hard work to distill these principles into standards, rules, and therefore into practice.
原则主义并非旨在作为一种算法或清单，以产生明确或可自动化的决策。相反，原则旨在产生张力，这种张力为裁决艰难的决策提供了共同的词汇和准则。这种共同的语言和共同的价值观发挥着强大的社会功能：确保社区成员（例如公司员工或产品用户）感到决策至少是合法的，并且经过了健康的流程，即使结果并非所有人都同意。
Principlism is not intended as an algorithm or checklist, yielding a clear or automatable decision. Instead, the principles are meant to be in tension, a productive tension that provides a common vocabulary and rubric for adjudicating difficult decisions. This common language and common set of values serves a powerful social function: ensuring that members of a community, such as the employees of a company or the users of a product, feel that the decision was at least made legitimately and with a healthy process, even if the result is not one with which everyone will agree.
贝尔蒙特的主要原则
Major Principles in Belmont
虽然贝尔蒙特委员会确定的原则借鉴了几个世纪的道德哲学，但委员们认为这些原则及其所有矛盾都存在于现有的社会规范中。在他们看来，“国家委员会几乎肯定认为这些原则已经植根于现有的公共道德中。” 17
While the principles identified by the Belmont commission drew on centuries of ethical philosophy, the commissioners took the principles, with all their tensions, to be present in existing social norms. In their view, “the national commission almost certainly believed that these principles are already embedded in preexisting public morality.”17
贝尔蒙特委员会的核心关注点是如何平衡科学实验可能带来的集体利益与对每个研究对象的影响。委员会的报告旨在捕捉合法目的与手段之间的矛盾，这体现在“尊重人”和“仁慈”这两个首要原则中。
A central concern of the Belmont commissioners was how to balance the collective good that might come from a scientific experiment with the impact on each of the individual research subjects. The commission’s report was designed to capture the tensions between legitimate ends and means, enshrined in the two first principles of “respect for persons” and “beneficence.”
尊重人意味着尊重作为研究对象参与研究的个人的自主权和尊严。这一原则通常被称为“知情同意”，源自哲学伦理学中的义务论传统，与伊曼纽尔·康德有着密切的联系。在人类受试者研究的背景下，这要求确保那些自主权被削弱的人（如儿童或被监禁者）的知情同意。
Respect for persons requires respecting the autonomy and dignity of individuals participating as research subjects. Often instrumentalized as “informed consent,” the principle derives from the deontological tradition within the philosophical ethics, strongly associated with Immanuel Kant. In the context of human subjects research this demands ensuring the informed consent of those with diminished autonomy, such as children or the incarcerated.
仁慈包括权衡研究项目的潜在利益和危害。这通常被概括为“不造成伤害”，但更一般地，这是指不仅对研究对象而且对社会，最大限度地提高利益并尽量减少伤害。最近，这一原则已扩展到人类社会以外的危害，例如对其他生物或环境的危害。这一原则本身源自结果主义或功利主义哲学传统，与约翰·斯图尔特·密尔、杰里米·边沁等人有关。
Beneficence includes to weighing the potential benefits and harms of a research project. Often this is summarized as “Do no harm,” but more generally this refers to maximizing benefit and minimizing harm not only to research subjects but to society. More recently this principle has extended to harms beyond human society, for example, to other living creatures or to the environment. This principle itself derives from the consequentialist or utilitarian philosophical tradition, associated with John Stuart Mill, Jeremy Bentham, and others.
这一原则尤其受到算法伦理的挑战，因为复杂的算法使得推测可能的意外影响和潜在危害变得困难。另一方面，算法产品和服务（如推荐引擎）也使得监测和减轻此类危害成为可能。与对于必须召回和修理的有缺陷的产品，可以调整算法并以数字方式重新部署。
This principle is particularly challenged by algorithmic ethics, in that complex algorithms make speculating on the possible unintended effects and potential harms difficult. On the other hand, algorithmic products and services like recommendation engines also make possible the monitoring and mitigation of such harms as they are revealed. Unlike a defective product which must be recalled and repaired, an algorithm can be tuned and digitally redeployed.
贝尔蒙特的第三项原则是正义，重点不是目的与手段之间的矛盾，而是公平的规范。特别是在对被监禁者的研究背景下，委员们不仅关注平等待遇，还关注压迫和分配不均。几年后，即 2004 年，卡伦·勒巴克兹教授回顾了她在委员会中的角色，强调了委员会对正义的承诺。她解释说，现在可以用更强烈的语言来表达这一承诺。“我们谈论正义，我们主要用平等待遇和保护弱势群体的语言来谈论它。一种我们当时没有使用但后来变得非常突出、对我来说非常重要的语言是压迫的语言。”她强调，这种说话方式将更清楚地表明对研究中的正义的承诺将意味着什么。“我认为，弱势群体和被压迫群体之间存在差异。正义需要纠正压迫，这可能会建立一些与我们多年前不同的结构。” 18
The third principle of Belmont is justice, focused not on the tension between ends and means but on norms of fairness. Particularly in the context of research on the incarcerated, commissioners were concerned not only with equal treatment but oppression and maldistribution. Reflecting upon her role on the commission some years later, in 2004, Professor Karen Lebacqz underscored the commission’s commitment to justice. She explained this commitment could now be cast in a stronger idiom. “We talked about justice and we talked about it primarily in the language of equal treatment and protection of the vulnerable. A language that we did not use in those days but that has become very prominent since and very important to me, is the language of oppression.” She underscored this way of speaking would bring out more clearly what a commitment to justice in research would entail. “I think there is a difference between populations who are simply vulnerable and populations who are oppressed. And, justice requires rectification of oppression and that might set some structures differently than the way that we did so many years ago.”18
这三项一般原则被认为暗含了额外的道德标准。例如，隐私可以被视为知情同意的一个例子——隐私被理解为披露事实的情况，而不是事实本身。例如，我们可能同意与医生分享一个事实，而我们不会与老师或学生分享。同样，“公平”被视为正义的一个基本方面。公平的目的是避免对穷人和被剥夺权利的人进行医学实验，这些人会承受这些实验的风险，而收益则流向有能力负担由此产生的药物或医疗费用的有权者。
The three general principles are taken to imply additional ethical standards. Privacy, for example, can be viewed as an example of informed consent—where privacy is understood as circumstances around a disclosure of the fact, rather than the fact itself. For example, we may consent to share a fact with a doctor that we would not share with our teachers or students. Similarly, “fairness” is viewed as a fundamental aspect of justice. Fairness aims to avoid, for example, medical experiments on the poor and disenfranchised, who suffer the risks of these experiments, whereas the benefits flow toward the empowered who can afford the resulting medicines or medical treatments.
这三项原则旨在“全面”涵盖人类受试者研究的应用伦理问题，但应用伦理意味着执行权力的转移。除了哲学工作外，委员们还提议将机构审查委员会程序编入法律——这是组织内部伦理的有力制度化。
These three principles were designed to be “comprehensive” for covering applied ethical problems of human subjects research, but applying ethics means a shift in power to enforce. Alongside their philosophical work, the commissioners proposed a codification of the Institutional Review Board process in law—a potent institutionalization of ethics within organizations.
从IRB到硅谷的制度化原则主义
Institutionalizing Principlism from IRB to Silicon Valley
除了基本原则之外，该委员会还单独发布了一份 132 页的提案，旨在制定流程设计来将这些原则付诸实施。19这些指导方针促成了机构审查委员会的成立，该委员会将作为联邦资助的条件来管理所有美国大学的人体研究。联邦资金只有在经过此类委员会的审查后才能用于项目，而该委员会的审议也应遵循这些共同的原则。控制资金赋予了机构审查委员会制定规则的权力，而这种权力是单纯的法规所不具备的。然而，这些委员会有多么不完善——毫无疑问，机构审查委员会的历史充满了不完善之处——其目的是为了确保研究人员按照原则行事。在其后的几年里，机构审查委员会不得不将这些原则应用于新的技术课题，包括基因工程和最近的计算社会科学研究。
Along with the foundational principles, the commission published—separately—a 132-page proposal for the creation of process design to operationalize them.19 These guidelines shaped the creation of the institutional review boards that would govern human subjects research at all US universities as a condition of federal funding. Federal funds may be directed to a project only after passing review by such a board, whose deliberations are guided by these shared principles. Controlling funds gives IRBs rules power, power that mere regulations might never have. However imperfectly— and there is no question the story of IRBs is replete with imperfections—these boards are designed to ensure that researchers act in accord with the principles. In the intervening years, IRBs have had to apply these principles to new technical topics, including genetic engineering and more recently research in computational social science.
IRB 模型继续充当着伦理制度化的主要模型。虽然我们可能不同意当代 IRB 系统的道德框架、规则或制度化，但关键在于该系统汇集了丰富的伦理哲学论述、在细微案例中将这种哲学反思付诸实践的方法，以及执行伦理的制度手段。框架。道德本身是没有约束力的；没有道德的法规只不过是官僚主义。
The IRB model continues to serve as a major model for institutionalizing ethics. While we may disagree with the ethical framework, the rules, or the institutionalization of the contemporary IRB system, the key point is that the system brings together a rich philosophical account of ethics, a means of putting that philosophical reflection into practice in nuanced cases, and institutional means for enforcing the framework. Ethics by itself is toothless; regulations without ethics are mere bureaucracy.
设计是“在一系列约束条件下对问题有意采取的解决方案”。20 IRB是流程设计的一个例子。当然，所有设计的例子也都是权力的陈述：尊重谁的意图，谁解决问题，谁设定约束。最终的设计，无论是产品还是流程，也会通过重新安排谁能对谁做什么来影响权力。
Design is “the intentional solution to a problem within a set of constraints.”20 The IRB is an example of process design. All examples of design, of course, are also statements of power: whose intent is respected, who solves the problem, and who sets the constraints. The resulting design, whether a product or a process, also impacts power by rearranging who can do what to whom.
随着社会科学家、活动家、计算机科学家和记者越来越多地开始指出大规模自动化决策系统（实践中的机器学习）对社区和民主的潜在和实际危险，贝尔蒙特原则和 IRB 结构为任何希望衡量技术影响并将其组织起来以实现利润以外的目的的人提供了一个强大的现有系统。对于像 Facebook 这样的公司，它们提供了一个自我监管的框架，而不是政府监管。正如 Facebook 的“发展 IRB”案例所表明的那样，旨在采用这些原则的公司面临的第一个挑战是重新语境化：这些在人类受试者研究中发展起来的原则如何应用于信息平台公司？旨在采用这些原则的公司面临的第二个挑战是决策的设计和权力分配：公司如何制度化组织设计和流程设计，以便这些原则有意义地约束和指导决策？
As social scientists, activists, computer scientists, and journalists increasingly began to signal the potential and realized dangers of large-scale automated decision systems—machine learning in practice—to communities and to democracy, the Belmont principles and the IRB structure offered a powerful existing system for anyone looking to gauge the impact of technologies and organize them toward ends other than, say, profit. For companies like Facebook, they offered a framework for self-regulation, rather than government regulation. As the case of Face-book’s “evolving the IRB” suggests, the first challenge for companies aiming to adopt these principles is one of recontextualizing: How do these principles, developed in the case of human subjects research, apply to information platform companies? The second challenge for companies aiming to adopt these principles is one of design and the distribution of power around decision-making: How do companies institutionalize organizational design and process design such that these principles meaningfully constrain and guide decisions?
事实证明，改进 IRB 并非易事，也并非有效。尤其是，将道德问题集中到 Facebook 的一个“拥有”道德的团体中，并没有阻止自呼吁数据道德管理这些公司以来日益高涨的呼声，因为数据赋权的算法的影响最大。过去十年来，道德规范以“原则”为框架，但这并不意味着数据驱动的平台公司会转型，通过聘请哲学家来组织业务。目前也不清楚这样做是否会产生任何实际效果。原则主义假定这些原则可能相互矛盾，需要以这些原则为共同词汇和价值观的个人进行善意裁决。虽然共同的道德原则植根于哲学传统，但雅各布·梅特卡夫、伊曼纽尔·莫斯和达娜·博伊德从一项针对平台公司员工的民族志研究中得出结论，应该更好地将道德理解为“社会现象，而不是主要的哲学抽象概念”。21实现道德过程需要组织成员的广泛认同，并授权组织遵守应用于其商业实践的原则。由于脱离了 IRB 的结构及其作为控制研究资金的“暂停点”的核心作用，目前尚不清楚如何说服同事重视道德原则，以及如何设计组织和流程，使这些原则可以作为约束——特别是如果这些约束会减少利润。
Evolving the IRB has proven not to be straightforward or effective. In particular, the centralization of ethical concerns into one group who “owns” ethics at Facebook has not stemmed the growing chorus in the years since calling for data ethics governing such companies, where data-empowered algorithms have the greatest impact. The framing of ethics over the past decade in terms of “principles” has not meant that data-powered platform companies pivoted, organizing their businesses by hiring philosophers. Nor is it clear that doing so would have any practical effect. Principlism presumes that the principles may be in tension with each other, requiring good faith adjudication by individuals who share the principles as their common vocabulary and values. While common ethical principles are grounded in a philosophical tradition, Jacob Metcalf, Emanuel Moss, and danah boyd conclude from an ethnographic study of employees at platform companies that ethics should be better understood “as social phenomena and not as primarily philosophical abstractions.”21 The realization of ethical process requires broad buy-in by the members of the organization and the empowerment to make organizations comport with principles applied to their business practices. Unmoored to the structure of the IRB and its central role as a “pause point” in controlling research funding, it is unclear how to convince colleagues to value ethical principles and how to design organizations and processes within which the principles could serve as any constraint—particularly if those constraints would reduce profit.
至少从某个方面来看，Facebook 的做法取得了巨大成功。牛津大学研究员 Brent Mittelstadt 在调查 2019 年人工智能伦理状况时发现，“人工智能伦理与医学伦理原则趋同……这是历史上最突出、研究最深入的应用伦理方法。”这对 Facebook 来说也是一种成功的方法，因为它“为政策制定者提供了不推行新法规的理由。” 22
In at least one respect, Facebook’s approach turned out to be a great success. Surveying the state of AI ethics in 2019, Oxford researcher Brent Mittelstadt found a “convergence of AI Ethics around principles of medical ethics . . . historically the most prominent and well-studied approach to applied ethics.” It was also a successful approach for Facebook in that it “provide[d] policy-makers with a reason not to pursue new regulation.”22
随着“不道德”行为的挑战逐渐引起社会科学家和媒体的关注，学者们开始质疑道德是否是道德的基石。贝尔蒙特原则及其随后的管理人类受试者研究的制度设计可以提供任何保护，以防止企业使用个人数据而产生的日益严重的危害和不公正。
As the challenges of “unethical” behavior came to draw attention from social scientists and, increasingly, the press, scholars questioned whether ethics, often in bedrock laid by the Belmont principles, and their ensuing institutional design for governing human subjects research could provide any shield against the rising harms and injustices of corporate uses of personal data.
拥有道德：流程、组织和权力
Owning Ethics: Process, Organization, and Power
除了呼吁道德原则之外，过去几年，人们越来越多地呼吁阐明公司内部有意义的道德流程可能包含哪些内容。例如，Inioluwa Deborah Raji 和合著者（包括 Margaret Mitchell、Timnit Gebru，我们在本章的开头也与他们合作）主张建立算法审计流程。23此前，当时任职于马库拉应用伦理中心的哲学家 Shannon Vallor 在商业伦理文献的基础上，开发了一套数字产品开发的检查点，在产品开发的不同阶段会提出不同的问题。24这些流程（审计和检查点）将道德质询与决策时刻结合起来，超过决策时刻，影响（包括危害）的规模就会增加。这些决策时刻也可以被视为行使权力的时刻——道德获得“牙齿”的时刻。它们超越了将公司视为一个统一的整体。如果道德决策与公司个人的目标（例如促销或成功的产品发布）之间没有紧密结合，那么道德如何成为社会不可或缺的一部分就不清楚。
Moving beyond calls for ethical principles, the past few years have seen a widening call for spelling out what a meaningful ethical process within a corporation might entail. As one example, Inioluwa Deborah Raji and coauthors, including Margaret Mitchell and Timnit Gebru, with whom we began this chapter, argue for a process for algorithmic auditing.23 Earlier, philosopher Shannon Vallor, then of the Markkula Center of Applied Ethics, building on the business ethics literature, developed a set of checkpoints in the development of a digital product, in which different questions are asked at different points in the product development.24 These processes—audits and checkpoints—couple ethical interrogation to moments of decision beyond which the scale of impact (including harms) increases. These decision moments can also be framed as moments where power is exercised—where ethics is given “teeth.” And they go beyond considering the firm as a uniform whole. Without strong coupling between ethical decision making and the objectives of individuals in corporations—such as promotions or successful product launches—it is unclear how ethics becomes integral to the community.
复杂的公司必然由具有不同利益的独立团队组成。这些利益可能包括“所有权”，即对特定收入流或用户行为的责任。通常不清楚在组织中应该将负责某项业务的团队设在何处，在公司中则不然。用语来说就是“拥有”道德。25确保个人将审计、暂停点或其他道德检查融入实践需要在公司内共享原则和不同参与者的激励之间取得一致。引用一位匿名技术员工的话：“你创建的系统必须是人们认为可以增加价值的东西，而不是一个没有增加价值的巨大障碍，因为如果它是一个没有价值的障碍，人们就不会去做，因为他们不需要。” 26这种成功的协调就是商业伦理学者西奥多·珀塞尔 (Theodore Purcell) 和詹姆斯·韦伯 (James Weber) 在 1979 年所描述的“道德的制度化……将道德正式而明确地融入日常商业生活……融入公司政策制定……融入所有日常决策和工作实践，融入所有就业层面。” 27
Complex companies necessarily comprise separate teams with separate interests. These interests may include “ownership,” meaning responsibility for a particular revenue stream or a user behavior. It is often unclear where in an organization to locate a team responsible for, in corporate parlance, “owning” ethics.25 Ensuring individuals will integrate audits, pause points, or other ethical checks into practice requires an alignment between shared principles and the incentives of diverse actors within a firm. To quote an anonymous tech employee, “the system that you create has to be something that people feel adds value and is not a massive roadblock that adds no value, because if it is a roadblock that has no value, people literally won’t do it, because they don’t have to.”26 This successful alignment is what business ethics scholars Theodore Purcell and James Weber described in 1979 as “[t]he institutionalization of ethics . . . getting ethics formally and explicitly into daily business life . . . into company policy formation . . . into all daily decision making and work practices down the line, at all levels of employment.”27
谷歌的案例凸显了愿望与实施之间的差距。甚至在谷歌道德 AI 团队进行备受瞩目的重组之前，谷歌在将自己定位为一家道德公司方面就已经遇到了麻烦。例如，该公司于 2019 年 3 月成立了一个外部委员会，就人工智能的道德影响提供建议，然后在内部和公众对其组成以及其与公司决策过程的整合提出批评后，于次月匆忙解散了该委员会。前谷歌员工梅雷迪斯·惠特克 (Meredith Whittaker) 在 2018 年将这些举动斥为“道德剧场”，并问道：“他们可以取消产品决策吗？否则他们有否决权吗？” 28
The case of Google underscores the gap between aspiration and implementation. Even before the high-visibility reorganization at Google’s Ethical AI team, Google had earlier stumbled at positioning itself as an ethical company. For example, the company had created an external council to advise on ethical implications of AI in March 2019, then hastily disbanded it the next month after internal and public criticism about its makeup as well as its integration into the company’s decision-making processes. Former Google employee Meredith Whittaker dismissed these moves as “ethics theater” in 2018, asking “Can they cancel a product decision? Do they have veto power otherwise?”28
谷歌并非个例。律师本·瓦格纳 (Ben Wagner) 指责科技公司“洗白道德”——试图逃避监管，却没有对道德做出有意义的定义，也没有设计流程来指导使用道德的决策。在《道德作为逃避监管的出路：从洗白道德到购物道德？》一文中，瓦格纳提出了公司道德流程的六项标准：
Google is not alone. The lawyer Ben Wagner accused tech companies of “ethics washing”—working to avoid regulation without meaningfully defining ethics and designing processes to guide decisions using ethics. In his “Ethics as an Escape from Regulation: From Ethics-Washing to Ethics-Shopping?,” Wagner proposes six criteria for ethical processes in companies:
1.相关利益者外部参与。
1. External participation of relevant stakeholders.
2. 外部和独立监督。
2. External and independent oversight.
3.决策过程透明。
3. Transparent decision-making process.
4.稳定的标准、价值观和权利清单。
4. A stable list of standards, values, and rights.
5.确保道德不会取代基本权利或人权。
5. Ensure that ethics do not substitute [for] fundamental rights or human rights.
6.“明确说明所作承诺与现有法律或监管框架之间的关系，特别是当两者发生冲突时如何处理。” 29
6. “A clear statement on the relationship between the commitments made and existing legal or regulatory frameworks, in particular on what happens when the two are in conflict.”29
最后一条标准将伦理与法律进行了对比：法律传统延续数千年，塑造了政府的运作和合法性，而应用伦理则旨在就决策合法性（尤其是当权者所作决策）在利益相关者之间形成共识。权力越来越多地掌握在数据赋能、活跃于国际的科技公司手中；利益相关者包括世界各地的公民，以及寻求制定法规以应对这种权力转移的国家领导人。
The final criterion contrasts ethics with law: while legal tradition stretches for thousands of years and shapes the process and legitimacy of government, applied ethics aims to form consensus among stakeholders as to the legitimacy of decisions, particularly those made by those in power. Increasingly, power is in the hands of data-empowered, internationally active technology companies; the stakeholders include citizens across the world, as well as leaders of state searching to craft regulation to respond to this shift in power.
正如我们将在最后一章中探讨的那样，倡导道德的个体员工拥有许多“人力”工具。他们经常发现，道德实践与雇主的财务目标存在严重冲突。尽管如此，企业对道德的外表越来越感兴趣。正如梅特卡夫和同事所写，“道德可以说是当今硅谷炒作周期中最热门的产品。” 30这在很大程度上是对监管和“国家权力”加强的威胁以及员工内部批评等“人力”的反应。
As we will explore in our final chapter, individual employees advocating for ethics have many “people power” tools at their disposal. Often, they find that ethical practices sharply conflict with the financial goals of their employers. Nonetheless, corporations are increasingly interested in the appearance of ethics. As Metcalf and colleagues write, “Ethics is arguably the hottest product in Silicon Valley’s hype cycle today.”30 Much of this is in reaction to threats of increased regulation and “state power,” as well as “people power” such as internal critiques by their employees.
“技术修复”的局限性
The Limits of “Tech Fixes”
如果道德问题可以通过技术手段解决，而不是通过复杂的社会协商行动，那会怎样？技术界的许多人都在寻求这样的解决方案。在希望推进算法伦理的技术人员群体中，公平和隐私这两个特定方面在过去几十年中已成为技术研究的重点领域。
What if ethical issues could be addressed through technological fixes, rather than complex social, deliberative action? Many in the technical community have sought just such solutions. Within the community of technologists hoping to advance algorithmic ethics, two particular facets—fairness and privacy—have blossomed into areas of technical research in the past decades.
麻省理工学院的大多数计算机科学研究生都不会在《法律、医学与伦理学杂志》上发表文章。Latanya Sweeney 显然不是受传统束缚的人。她担心仅仅从公开发布的数据库中删除“姓名”字段会产生匿名的假象，因此在一系列论文中令人信服地说明了如何将一个由于姓名被删除而被视为“匿名”的数据库与具有其他唯一标识符的第二个数据库相结合，重新识别个人，从而暴露敏感信息。同行评议论文可能是学术界的流行词，但几乎不具备现实世界实验的影响：为此，Sweeney 通过使用她所在州州长自己的“匿名”医疗记录重新识别他来说明这一点，该记录是通过结合公开投票记录发现的。31两者共有的标识符（出生日期、性别和邮政编码）共同构成了这把锁的钥匙。几年后，阿尔温德·纳拉亚南 (Arvind Narayanan) 和维塔利·什马蒂科夫 (Vitaly Shmatikov) 也展示了如何使用来自另一个数据库的数据对 Netflix Prize 数据集中的至少部分评论者进行去匿名化，从而导致 Netflix 撤回该数据集。32这种去匿名化可能会泄露高度个人化的偏好，这可能会让用户感到尴尬甚至危及安全。
Most MIT graduate students in computer science don’t publish in The Journal of Law, Medicine & Ethics. Latanya Sweeney was clearly not one to be constrained by convention. Concerned about the illusion of anonymity provided by merely removing the “name” field from publicly released databases, she convincingly illustrated in a series of papers how a database considered “anonymous,” thanks to having names redacted, could be combined with a second database with other unique identifiers to reidentify individuals and thus expose sensitive information. Peer-reviewed papers may be coin of the realm in academia, but don’t nearly have the impact of a real-world experiment: for this, Sweeney illustrated the point by reidentifying her state’s governor using his own “anonymous” medical record, revealed by combining with public voting records.31 The identifiers common to both (birthdate, gender, and zip code) together formed the key to this lock. A few years later, Arvind Narayanan and Vitaly Shmatikov similarly showed how to deanonymize at least some of the reviewers in the Netflix Prize data set using data from another database, causing Netflix to pull the data set.32 Such deanonymization threatened to reveal highly personal preferences that could embarrass and even endanger users.
Sweeney 提出了一种技术防御措施来应对此类攻击：k-匿名性，即数据库的一种属性，其中没有记录是唯一的，但与至少 k-1 个其他此类记录相同。33例如，在投票记录中，我们可以只发布出生月份（而不是日期），或者只发布邮政编码的前三位或四位数字（而不是全部五位），直到我们确保任何一条记录都无法被唯一识别。直观地讲，这样的过程提供了一定程度的合理否认：“这不是我；这是这个数据库中其他 k-1 条相同记录之一！”
Sweeney proposed a technical defense against such an attack: k-anonymity, an attribute of a database in which no record is unique, but is identical to at least k-1 other such records.33 For example, in the voting record we could have released only birth month (rather than date), or only the first three or four digits of a zip code (rather than all five), until we ensure that any one record cannot be uniquely identified. Intuitively, such a process provides a level of plausible deniability: “It wasn’t me; it was one of the k-1 other identical records in this database!”
类似技术上可行的否认形式是差分隐私，这是一种提供隐私的随机方法。部分受到“斯威尼实现的惊人隐私妥协”的启发，辛西娅·德沃克于 2006 年提出了差分隐私，作为一种噪声生成技术，这样原始数据库就永远不会被泄露。34与k 匿名一样，我们必须选择所需的粒度k，差分隐私也带有直接的主观设计选择，例如要注入的噪声强度以及噪声模型本身（例如，如果我们要在数据库中查询文档中包含的单词而不是患者的身高，我们会选择不同的噪声注入数学模型）。这种粒度的选择说明了隐私和效用之间的紧张关系。正如德沃克在提出该技术的原始著作中所写，隐私“需要一些效用的概念——毕竟，一种始终输出空字符串或纯随机字符串的机制显然可以保护隐私。” 35由于美国人口普查局决定在发布 2020 年人口普查记录时使用差异隐私，差异隐私在过去几年中不断得到完善、发展和扩展，并引起了广泛关注。
A similar technically plausible form of deniability is that of differential privacy, an aleatory approach to providing privacy. Motivated in part by “the spectacular privacy compromises achieved by Sweeney,” Cynthia Dwork proposed differential privacy in 2006 as a noise-generating technique, such that the original database is never revealed.34 As with k-anonymity, in which we must choose the granularity k desired, differential privacy comes with immediate subjective design choices as to the strength of the noise to be injected, as well as the noise model itself (e.g., if we are to query a database for words contained in a document rather than heights of patients, we would choose a different noise-injecting mathematical model). Such a choice of granularity illustrates the tension between privacy and utility. As Dwork wrote in the original work proposing the technique, privacy “requires some notion of utility—after all, a mechanism that always outputs the empty string, or a purely random string, clearly preserves privacy.”35 Differential privacy has continued to be refined, developed, and extended, with a flurry of attention over the past years thanks to the decision by the US Census Bureau to use differential privacy when releasing records from the 2020 census.
在我们的第一章中，我们通过她在 2014 年“机器学习中的公平性、责任感和透明度”（最初为 FAT-ML，现为 FAccT）研讨会上的演讲认识了计算机科学研究员 Hanna Wallach。在接下来的几年里，随着这个社区的发展，更多来自计算机科学领域的技术导向型从业者专注于开发高度数学和技术性的方法来定义和量化公平性——更多的是代码，更少的哲学和法律。在这些技术文献蓬勃发展的同时，关于算法危险性的文献也日益增多，其中包括 Cathy O'Neil、Virginia Eubanks 和 Ruha Benjamin 等人的早期著作。36虽然公平的概念几十年来一直是美国法律的一部分（特别是在 1964 年《民权法案》之后），但直到最近，公平性在技术文献中才略有体现。随着这些文献的迅速扩展，一些特别的意外因素塑造了将工程思维应用于公平性问题的目标。首先，公平性有许多合理的量化定义；其次，其中一些定义是互相不相容的——无论是在形式上还是在实践中。
In our first chapter we met the computer science researcher Hanna Wallach via her 2014 talk at the workshop named “Fairness, Accountability, and Transparency in Machine Learning” (originally FAT-ML, now FAccT). Over the course of the next few years, as this community grew, more technically oriented practitioners from computer science focused on developing highly mathematical and technical approaches on defining and quantifying fairness—more code, less philosophy and law. In parallel to the blossoming of this technical literature, a literature on the dangers of algorithms grew, including germinal works by Cathy O’Neil, Virginia Eubanks, and Ruha Benjamin, among others.36 While notions of fairness have been part of US law for decades (particularly after the Civil Rights Act of 1964), fairness had figured only slightly in the technical literature until recently. As this literature rapidly expanded, some particular surprises shaped the goal of applying engineering mindset to problems of fairness. The first was that there are many plausible quantitative definitions of fairness; the second was that some of these definitions are mutually incompatible—both formally and in practice.
为了说明量化公平的挑战，我们来看看 2016 年 5 月的“机器偏见”案例。37非营利新闻机构 ProPublica 的一篇文章调查了 COMPAS，这是 Northpointe 公司开发的专有算法，用于预测佛罗里达州布劳沃德县的犯罪累犯率。仔细的研究表明，该算法是不公平的，因为被算法评为“高犯罪率”的白人被告实际上比被评为“高犯罪率”的黑人被告更有可能继续犯罪。然而，仅仅两个月后，Northpointe 的三位研究人员就发表了自己的分析，表明其方法是公平的，因为该算法“对黑人和白人同样准确”。普林斯顿计算机科学家 Arvind Narayanan 用 21 种不同的公平解释说明了不同的技术定义如何具有截然不同的政治意义。38除了 Narayanan，Solon Barocas 和 Moritz Hardt ，最初的 FATML 研讨会的两位联合组织者总结了三项核心公平措施，因为它们可能适用于种族歧视问题：
To illustrate the challenge of quantifying fairness, consider the case of “Machine Bias” from May 2016.37 The piece by the journalism nonprofit ProPublica investigated COMPAS, a proprietary algorithm developed by the company Northpointe to predict criminal recidivism in Broward County, Florida. Careful work showed that the algorithm was unfair in the sense that the group of white defendants who were algorithmically scored to be “high crime” were in fact more likely to go on to commit crimes than the similarly scored Black defendants. However, three researchers at Northpointe published their own analysis just two months later, showing that its methods were fair in the sense that the algorithm is “equally accurate for blacks and whites.” Princeton computer scientist Arvind Narayanan has illustrated how different technical definitions have radically different politics, using twenty-one alternate accounts of fairness.38 Along with Narayanan, Solon Barocas and Moritz Hardt, two of the co-organizers of the original FATML workshop, summarize three central fairness measures, as they might be applied to questions of racial discriminations:
独立性：模型输出与种族无关（用高尔顿十九世纪的话来说就是“不相关”）。
Independence: the model output is independent of (in the nineteenth-century language of Galton, “uncorrelated with”) race
分离：考虑到真实结果（例如，当被告实际上犯下或未犯下后续罪行时，被视为单独的群体），算法的得分与他们的种族无关
Separation: given the true outcome (for example, when defendants who in fact did or did not commit later crimes are considered as separate groups), the algorithm’s score is independent of their race
充分性：给定算法的分数（例如，当预测被告会或不会犯下后续罪行时，将其视为单独的群体），真实结果与种族无关。
Sufficiency: given the algorithm’s score (for example, when defendants who were predicted to commit or not to commit later crimes are considered as separate groups), the true outcome is independent of race.
这些条件可以更用数学的方式表述，并且适用于受保护属性（即不一定是种族）和一般结果的一般情况。*
These conditions can be stated more mathematically, and for general cases of protected attributes (i.e., not necessarily race) and general outcomes.*
公平性定义的模糊性使得原本是机器学习的一条老路变得复杂：统计优化。正如我们在机器学习一章中看到的那样，机器学习中的计算方法现在已经被充分证明可以学习哪些策略可以优化期望的目标，即使在结合算法、数字产品和社会的复杂环境中也是如此。即使目标包括竞争，这些方法也可以发挥作用，例如，“度量”（例如统计准确性）和“反度量”（例如模型复杂度）。39计算机科学研究人员 Michael Kearns 和 Aaron Roth 在他们最近的著作《道德算法》中提倡这种方法。“从科学、监管、法律或道德角度来看，对这一事实唯一明智的回应是承认它，并试图直接衡量和管理准确性和公平性之间的权衡。” 40然而，技术修复只能到此为止。例如，即使是针对公平性和准确性进行优化的最佳算法也无法解决统计上自我强化的过度警务等问题，在这种警务中，“犯罪”和“逮捕”的预测被混为一谈，向之前观察到更多逮捕的地区派遣更多警察。41正如 Kearns 和 Roth 所观察到的，算法只是社会技术系统的一部分：“好的算法设计可以指定一系列解决方案，但人们仍然必须从中选择一个。” 42
The ambiguity over which definition of fairness to use complicates what would otherwise be a well-trodden path for machine learning: statistical optimization. As we saw in our chapter on machine learning, computational methods in machine learning are by now well proven at learning what policies can optimize a desired objective, even in complex environments combining algorithms, digital products, and societies. These methods can work even when the objective includes a competition, for example, between a “metric” (e.g., statistical accuracy) and a “countermetric” (e.g., model complexity).39 The computer science researchers Michael Kearns and Aaron Roth advocate this approach in their recent book The Ethical Algorithm. “The only sensible response to this fact—from a scientific, regulatory, legal, or moral perspective—is to acknowledge it and to try to directly measure and manage the trade-offs between accuracy and fairness.”40 Tech fixes, though, only go so far. For example, even an optimal algorithm—optimized for fairness and accuracy—will not fix problems such as statistically self-reinforcing over-policing, in which the prediction of “crime” and “arrests” are conflated, sending more police to an area in which more arrests have been previously observed.41 As Kearns and Roth observe, the algorithm is only one part of a socio-technical system: “Good algorithm design can specify a menu of solutions, but people still have to pick one of them.”42
即使这些技术方法奏效（而且通常确实奏效），它们也必然需要组织内部（或组织之上）的权力，需要有执行和指导的权力，而不仅仅是批评。正如谷歌道德人工智能团队的倒闭所表明的那样，在组织结构图中，尚不清楚可以将做出此类决定的机构放在哪个位置。
And even when these technical approaches work—and they often do—they necessarily require power in (or over) organization, with authority to enforce and direct, not merely to critique. As the collapse of the Ethical AI team at Google illustrates, it is unclear where, within the organizational chart, one could place the agency to make such decisions.
尽管这些技术解决方案非常重要，但它们的重点是重新设计算法系统和数据收集的各个方面，以最大限度地减少偏见和结构性不平等的影响，但并不能改变推动和维持不平等的社会结构。它们追求公平，而不是更坚定地追求正义。凯瑟琳·迪伊格纳齐奥和劳伦·克莱因写道：“更广泛地关注数据正义，而不仅仅是数据伦理，可以帮助确保过去的不平等不会被提炼成黑箱算法。” 43
For all their importance, such technical solutions focus on reworking facets of algorithmic systems and the collection of data to minimize bias and the effects of structural inequality, but do not function to alter social structures driving and maintaining the inequality. They strive for fairness, rather a more robust pursuit of justice. “A broader focus on data justice,” write Catherine D’Ignazio and Lauren Klein, “rather than data ethics alone, can help to ensure that past inequalities are not distilled into black boxed algorithms.”43
鉴于自动算法系统日益重要，我们社会的正义本身也越来越依赖于数据正义。在对伦理技术方法的批评中，萨菲亚·诺布尔 (Safiya Noble) 和马修·勒布伊 (Matthew Le Bui) 同样认为，“在这些权力体系面前，仅仅追求公平，并不能解决数字技术日益成为其他形式结构性权力的核心这一问题。” 44在制定人工智能伦理时，太多研究人员转向了贝尔蒙特报告中最程序化的方面，而忽视了其在社会经济、性别和种族差异面前对实质性正义的关注。
And given the growing centrality of automated algorithmic systems, justice in our societies itself depends increasingly on data justice. In their critique of technical approaches to ethics, Safiya Noble and Matthew Le Bui likewise argue, “Simply striving for fairness in the face of these systems of power does little to address the ways that digital technologies are increasingly central to other forms of structural power.”44 In crafting AI ethics, too many researchers turned to the most procedural facet of the Belmont report, while missing its concerns with substantive justice in the face of socioeconomic, sexual, and racial disparities.
最终目标永远是“修复”人工智能系统，而不是使用不同的系统或根本不使用系统。
The endgame is always to “fix” A.I. systems, never to use a different system or no system at all.
— 朱莉娅·波尔斯和海伦·尼森鲍姆45
—Julia Powles and Helen Nissenbaum45
人们渴望通过技术手段解决人工智能的问题，其前提是人工智能的使用将会得到改进，而不是被完全推迟甚至抵制。律师兼技术学者 Frank Pasquale 将质疑系统构建的运动称为算法问责的“第二波”浪潮：“第一波算法问责专注于改进现有系统，而第二波研究则质疑是否应该使用它们——如果应该，谁来管理它们。” 46越来越多的技术专家加入律师、社会学家和活动家的行列，提出这些更具结构性的问题，并采取行动重新安排权力，因为私人经常合作，敦促企业和政府实现更大的公平和正义。
The aspiration to apply a technical fix to problems in AI presumes that the use of AI is there to be improved, rather than pushed back or even resisted entirely. Lawyer and technology scholar Frank Pasquale identifies the movement to question even the building of systems as a “second wave” of algorithmic accountability: “While the first wave of algorithmic accountability focuses on improving existing systems, a second wave of research has asked whether they should be used at all—and, if so, who gets to govern them.”46 A growing number of technologists are joining lawyers, sociologists, and activists in posing these more structural questions and taking action to enact a reordering of power as private individuals often working together to press corporations and governments toward greater fairness and justice.
（自我）监管俘获
(Self-)Regulatory Capture
个人的隐私秩序越来越受到越来越多机构的关注，这些机构试图通过修复或挑战来解决算法中的伦理问题人工智能的使用。然而，许多这样的“自我监管”组织本身是由他们试图批评的公司资助的，从而导致冲突，而这种冲突可能会减缓、抑制和巧妙地引导这种批评。在《‘道德人工智能’的发明：大型科技公司如何操纵学术界以避免监管》一书中，现任莱顿大学教授、当时是麻省理工学院博士生的罗德里戈·奥奇加梅 (Rodrigo Ochigame) 追踪了资金从科技公司流向研究机构的过程，这些研究机构旨在创建一个“道德人工智能”领域，以批评和限制这些公司的盈利产品和服务。47人工智能研究人员和许多 (也许是大多数) 道德人工智能研究人员都发现自己深深地依赖于少数几家公司。几位澳大利亚学者最近认为，“诱使研究人员成为可以轻易解雇或纳入的美德供应商……对现有的商业组织形式或商业模式几乎没有抵抗力。” 48然而，正如我们之前所说，没有权力的道德可能是无效的，而没有道德的权力则缺乏任何积极的社会和政治方向。
The private ordering of individuals is increasingly resonant with a growing number of institutions which seek to address ethical concerns in algorithms, either via fixes or by challenging the use of AI. Many of these “self-regulatory” organizations are however themselves funded by the companies they seek to critique, leading to a conflict which can slow, stifle, and subtly direct such criticism. In “The Invention of ‘Ethical AI’: How Big Tech Manipulates Academia to Avoid Regulation,” Rodrigo Ochigame, now a professor at Leiden University but then a PhD candidate at the Massachusetts Institute of Technology, traced the flow of funding from technology companies to research institutions aiming to create a field of “Ethical AI” that would critique and constrain these companies’ profit-generating products and services.47 Both AI researchers and many (perhaps most) ethical AI researchers find themselves profoundly dependent on a small array of corporations. “Enticing researchers,” several Australian scholars have recently argued, “to become suppliers of virtue that can be easily dismissed or incorporated . . . offers little resistance to existing forms of business organisation or business models.”48 And yet, as we have suggested before, ethics without power may be inert, and power without ethics lacks any positive social and political direction.
算法产品的不透明性及其危害和影响，以及公司长期存在的组织复杂性，使得“践行道德规范”变得困难。难以量化的长期道德问题与短期量化问题之间的紧张关系只会加剧这些挑战——通常以围绕所谓“指标”优化的组织原则来表达。例如，隐私承诺可能会受到“监视资本主义”盈利能力的挑战，用 Shoshana Zuboff 倡导的一句话来说：对个人的加强跟踪以及对用于营销和其他领域的此类细粒度数据的经济需求。49作为一种技术，由此类数据驱动的算法构成了 Zeynep Tufekci 所说的“说服架构”——同样有效地使用无论说服是为了支持产品还是政治候选人。要理解这些架构的权力和盈利能力，我们必须从人类作为道德决策者的观点转向人类作为有价值关注的来源的观点。在这个领域，正如艺术家卡洛塔·费伊·斯库尔曼和理查德·塞拉在 1973 年所写的那样，“你就是产品。”
The opacity of algorithmic products and their harms and impacts, as well as long-standing organizational complexities at firms, makes “doing ethics” difficult. The challenges are only amplified by the tension between difficult to quantify long-term ethical concerns and short-term quantitative concerns—typically expressed as an organizing principle around optimization of so-called “metrics.” For example, commitments to privacy can be challenged by the profitability of “surveillance capitalism,” to use a phrase championed by Shoshana Zuboff: the enhanced tracking of individuals and the economic demand for such granular data for use in marketing and beyond.49 As a technology, the algorithms driven by such data constitute what Zeynep Tufekci terms “persuasion architectures”—used equally effectively whether the persuasion is in support of a product or a political candidate. To understand the power and profitability of these architectures, we must turn from the view of humans as ethical deciders to humans as sources of valuable attention. In this arena, as the artists Carlota Fay Schoolman and Richard Serra wrote in 1973, “you are the product.”
*有关这些更数学术语的定义，请参阅 Solon Barocas、Arvind Narayanan 和 Moritz Hardt 的《公平与机器学习》，2019 年， https://fairmlbook.org/。
* For these definitions in more mathematical terms, see Solon Barocas, Arvind Narayanan, and Moritz Hardt, “Fairness and Machine Learning,” 2019, https://fairmlbook.org/.
Persuasion, Ads, and Venture Capital
在信息丰富的世界里，信息的丰富意味着……信息消耗的东西稀缺。信息消耗的东西相当明显：它消耗了接收者的注意力。
In an information-rich world, the wealth of information means . . . a scarcity of whatever it is that information consumes. What information consumes is rather obvious: it consumes the attention of its recipients.
–赫伯特·西蒙，1971 1
–Herbert Simon, 19711
观看时间是首要任务……其他一切都被视为干扰。
Watch time was the priority. . . . Everything else was considered a distraction.
–[前] 谷歌工程师 Guillaume Chaslot 在 2018 年描述了 YouTube 推荐引擎的唯一 KPI 2
–[ex]-Google engineer Guillaume Chaslot, describing, in 2018, YouTube’s recommendation engine’s sole KPI2
1929 年 3 月 31 日星期日，复活节游行在第五大道举行， “人行道上五彩缤纷，灯火辉煌”，“现代、繁荣的纽约正在欢庆”。《纽约时报》头版报道称，“游行进行到高潮时，大约有十几名年轻女子在圣托马斯教堂和圣帕特里克教堂之间来回漫步，大摇大摆地抽着烟。其中一人解释说，香烟是‘自由的火炬’，照亮了女性像男性一样在街上随意吸烟的那一天。” 3《纽约时报》轻信地报道了这一事件，但实际上并不是一场女性吸烟者通过烟草产品推动性别平等的自发起义。这场表演由“公共关系之父”爱德华·伯内斯 (Edward Bernays) 在美国烟草公司的资助下策划。伯内斯进而创立了公共关系领域，并提倡“宣传”（这是他 1926 年出版的有关这一主题的书的名称，当时这个词还不完全是嘲讽用语），以确保民主的正常运转。4他在政治和营销领域之间无缝穿梭，敏锐地意识到“同意工程”的公开和隐蔽动态，他的远见卓识领先于时代一个世纪。在本章中，我们将追溯过去一个世纪将注意力货币化的努力，包括最近加入这一运动的商业权力。就像化肥和汽油一样，广告和风险投资 (VC) 投资本身可能看起来相当平庸；然而，正如我们将看到的，它们结合在一起会产生爆炸性的组合。
“The sidewalks were a bright medley of color” on Fifth Avenue at the Easter Parade on Sunday, March 31, 1929, as “modern, prosperous New York was celebrating.” According to the front page of The New York Times, “About a dozen young women strolled back and forth between St. Thomas’s and St. Patrick’s while the parade was at its peak, ostentatiously smoking cigarettes. One of the group explained that the cigarettes were ‘torches of freedom,’ lighting the way to the day when women would smoke on the street as casually as men.”3 What The Times reported on, credulously, was not in fact a spontaneous uprising of women smokers advancing gender equality via tobacco products. The performance had been engineered by Edward Bernays, “the father of public relations,” with funding from the American Tobacco Company. Bernays went on to create the field of public relations and advocate for Propaganda (as he titled his 1926 book on the subject, not yet exclusively a term of derision) to ensure functioning democracy.4 Passing seamlessly between the worlds of politics and marketing, and astutely aware of the overt and covert dynamics of “The Engineering of Consent,” his was a vision a century ahead of its time. In this chapter we’ll trace the rise of efforts over this past century to monetize attention, including the business forces which have recently been brought to bear on this drive. Like fertilizer and gasoline, advertising and venture capital (VC) investing might seem rather banal on their own; yet, as we will see, they give rise to an explosive combination when mixed.
注意力货币化
Monetizing Attention
正如题词所暗示的那样，经济学家和人工智能先驱赫伯特·西蒙很早就意识到了计算机和信息处理的兴起将带来什么：注意力经济。西蒙认为，由于计算机使信息的存储和传输几乎免费，注意力的稀缺性将产生更大的价值，从而出现人们注意力的经济。“仅仅知道生产和传输信息的成本是不够的；我们还必须知道，就稀缺的注意力而言，接收信息的成本是多少。” 5这场早期的数字洪流与现有的行业——广告业——正面交锋。在西蒙的预言发表后不久，艺术家理查德·塞拉和卡洛塔·费伊·斯库尔曼谴责了当时占主导地位的信息传输机制——电视，这种包罗万象、由企业赞助的世界观被包装、出售和强加给人们。塞拉解释说，他们努力“明确”广播电视的“资本主义现状”及其商业模式：免费服务提供给人们，以换取那些想要说服的人偶尔的打断。在他们的视频作品“电视传递人们”中，他们争论道
As the epigraph suggests, the economist and AI pioneer Herbert Simon realized early what the rise of computers and information processing would bring: an attention economy. Simon argued that as computers made the storage and transmission of information nearly free, greater scarcity of attention would accrue greater value, and thereby an economy for people’s attention would emerge. “It is not enough to know how much it costs to produce and transmit information; we must also know how much it costs, in terms of scarce attention, to receive it.”5 This early digital deluge met head-on with an existing industry: the business of advertising. Not long after Simon’s prophecy, the artists Richard Serra and Carlota Fay Schoolman decried the all-enveloping, corporate-sponsored worldview packaged, sold, and imposed on people by the dominant mechanism of information transmission of the time: television. Serra explained that they strove “to make explicit” the “capitalist status quo” of broadcast television and its business model: a free service is provided to people in exchange for occasional interruptions funded by those who wish to persuade. In their video piece “Television Delivers People,” they argued
商业电视每分钟可传送至2000万人。
Commercial television delivers 20 million people a minute.
在商业广播中，观众需要付费才能获得自我推销的权利。
In commercial broadcasting the viewer pays for the privilege of having himself sold.
被消费的是消费者。
It is the consumer who is consumed.
你是电视的产物
You are the product of t.v.
您被传送给作为客户的广告商。
You are delivered to the advertiser who is the customer.
他消耗了你。
He consumes you.
观众不对节目负责
The viewer is not responsible for programming
...
...
你就是最终产品。6
You are the end product.6
20 世纪 80 年代，后来担任纽约大学文化与传播系主任的尼尔·波兹曼 (Neil Postman) 就这种模式可能出现的问题提出了警告。7他阐明了这种广告模式如何扭曲了赞助商与内容制作者之间的关系，尤其是在 20 世纪 80 年代的广播电视中，扭曲了广告商与媒体之间的关系。波兹曼认为，媒体制作的内容更吸引眼球，而不是事实，因为实际吸引的观众数量会为广播公司带来更多收入。事实上，这带来了一种危险：内容创作者通常会被激励去创作同样有趣的内容尽可能地；他将自己的书命名为《娱乐至死》（1985 年）。8这种扭曲的负面影响有限，因为广播电视作为定义大众世界观的主要真理来源，其控制力有限（尽管很大）。
In the 1980s, Neil Postman, later the chair of NYU’s Department of Culture and Communication, offered a warning about how this model could go wrong.7 Namely, he illuminated how this advertising model perverts the relationship between the sponsor and the content producer, and specifically in the case of 1980s broadcast television, the relationship between the advertisers and media. Postman argued that the media creates content that is more attention-grabbing than factual, since the actual number of viewers grabbed leads to more money earned for the broadcaster. Indeed, this creates a danger: that content creators more generally are incentivized to create content which is as entertaining as possible; he entitled his book on the subject Amusing Ourselves to Death (1985).8 A limit to the negative impact of such a perversion was the finite (though large) stranglehold broadcast television had as the primary source of truth in defining the mass public worldview.
换句话说：损害的规模受到电视观众规模的限制。
Said otherwise: The scale of the damage was limited by the scale of television viewing public.
网络
The Net
万维网 (WWW) 项目旨在允许链接到任何地方的任何信息。...如果您有兴趣使用该代码，请给我发邮件。它非常原型化。...
The WorldWideWeb (WWW) project aims to allow links to be made to any information anywhere. . . . If you’re interested in using the code, mail me. It’s very prototype [sic] . . .
— 蒂姆·伯纳斯·李爵士，1991 年 8 月 6 日 14:56:20 GMT，在 alt.hypertext 9上宣布 WWW
—Sir Tim Berners-Lee, 6 Aug 91 14:56:20 GMT, announcing the WWW on alt.hypertext9
随着我们的信息来源（以及随之而来的广告经济）转移到万维网，这种动态——新兴的“注意力经济”（其中“你就是产品”）发生了怎样的变化？“网络”的诞生通常定在 1991 年 8 月 6 日，那天 Tim Berners Lee 在一个 Usenet 组上发布了关于“万维网项目”的帖子。到 1994 年，多家公司成立，只是为了试图组织日益复杂的网络，因为它在没有任何集中控制的情况下迅速发展，用户难以浏览。在线广告可以追溯到电子邮件（20 世纪 70 年代）和 Usenet 讨论组时代，甚至在网络出现之前。到 1996 年，几家公司开始销售在线“横幅广告”和干扰性“弹出窗口”。10到了 20 世纪 90 年代中期，大量广告业和“电子商务”（现在简称为“商业”）公司一起蓬勃发展，其中最引人注目的是 eBay 和亚马逊网站，这两家公司均成立于 1995 年，此外还有一大批在第一次“互联网泡沫”破灭时破产的公司。
How did this dynamic—the nascent “attention economy,” in which “you are the product”—change as our diet of information (and the economy of advertising along with it) moved to the World Wide Web? “The Web” is often given the birthday of August 6, 1991, when Tim Berners Lee posted to a Usenet group about “the WorldWideWeb project.” By 1994, multiple companies had formed simply to try to organize the spiraling complexity of the web, as it rapidly grew without any centralized control, becoming difficult for users to navigate. Online advertising dates from the time of email (the 1970s) and Usenet discussion groups, even before the web. By 1996, several companies were selling online “banner ads” and interruptive “pop-ups.”10 By the mid-1990s, abundant advertising flourished along with “e-commerce” (now simply called “commerce”) companies, notably including the sites eBay and Amazon, both founded in 1995 as well as a welter of companies that failed in the first “dot-com” bust.
在此背景下，前粒子物理学家、现媒体学者迈克尔·戈德哈伯 (Michael Goldhaber) 在新的互联网研究期刊《第一个星期一》 (First Monday，创刊于 1996 年)中更新了西蒙关于注意力稀缺的观察：
Against this backdrop, the former particle physicist turned media scholar Michael Goldhaber updated Simon’s observation about the scarcity of attention in a new internet studies journal called First Monday (formed in 1996):
然而，信息不可能成为经济的基础，原因很简单：经济是由稀缺资源所支配的，而信息，尤其是网络上的信息，不仅丰富，而且泛滥成灾。11
Information, however, would be an impossible basis for an economy, for one simple reason: economies are governed by what is scarce, and information, especially on the Net, is not only abundant, but overflowing.11
信息并不稀缺。因此，流动或受限制的东西是注意力。
Information is not scarce. And so, the thing that is flowing or that is constrained is attention.
戈德哈伯是在缺乏危机的特定时期写作的；他写道“有那么多闲暇时间，但我们都感到忙碌”，因为“我们所有的物质享受都得到了考虑。”（显然，他是从第一世界的富裕和舒适的角度来写作的。）“因此，我们有大量时间上网。” 12十年后，BuzzFeed 的创始人乔纳·佩雷蒂 (Jonah Peretti) 也做出了类似的观察；他将 BuzzFeed 的目标市场称为“工作无聊网络”，意思是数百万人的工作虽然可以上网，但却让他们感到无聊，因此注意力过剩：上网的时间。13 Buzzfeed最初用可爱的小猫和性感的名人照片来满足这种过剩的需求；它后来成为一家“独角兽”——一家估值超过 10 亿美元的初创公司。
Goldhaber was writing from a particular time of lack of crisis; he writes that “there’s so much leisure time and yet we all feel busy” because “all of our creature comforts are taken into account.” (Obviously, he was writing from a First World perspective of affluence and comfort.) “And so, we have lots of time to surf the web.”12 A similar observation was made a decade later by Jonah Peretti, founder of BuzzFeed; Buzz-Feed’s target market he called the “Bored at Work Network,” meaning the millions of people at jobs which provided ‘Net access but which left them bored and thus with a surfeit of attention: time to surf.13 Buzzfeed met this glut, initially, with cute kittens and racy celebrity pictures; it would come to be a “unicorn”—a start-up company valued at over a billion dollars.
戈德哈伯意识到，虽然注意力已经很有价值，但网络让任何人都可以消耗其他人的（在线）注意力。伯内斯 1929 年复活节的发明需要大量的协调，并且需要有可以利用的吸引注意力的游行；有了网络，任何人都可以发布信息，从而有可能消耗在线上任何其他人的注意力。戈德哈伯认为，这种变化将增加品牌的重要性个人相对于个人雇主或任何公司的影响力。就像明星记者可能会离开报纸去建立自己的博客，或者后来建立自己的时事通讯（或者最近建立自己的 Substack）一样，网络允许个人在没有资金充足或强大的现任者支持的情况下发展自己对在线关注的掌控。14
Goldhaber realized that while attention was already valuable, the web made it possible for anyone to consume anyone else’s (online) attention. Bernays’s 1929 Easter Day contrivance required tremendous coordination and the existence of an attention-getting parade which could be exploited; with the web, anyone can publish and thereby potentially consume the attention of anyone else online. Goldhaber argued that this change would increase the importance of the brand of the individual relative to that of the individual’s employer, or any corporation. In the same way that star journalists might leave a newspaper to establish their own blog, or later their own newsletter (or even, more recently, their own Substack), the web allows individuals to develop their own command of online attention without the backing of a well-financed or powerful incumbent.14
反思一下网络出现之前经济与注意力之间的关系有多么不同是很有用的。1997 年，关于电子书，戈德哈伯写道：“目前，直接通过互联网分发书籍是不切实际的，尽管很容易预见到这种情况不会持续很长时间，而且实体书将被视为笨重和古怪的。”事实上，在 COVID-19 大流行期间，学者对书籍和图书馆的依赖让位于电子书和扫描件，帮助撰写您手中的书（或您手掌中的电子书）——然而，纸质书的销量却激增。同样，付费墙规范之间的关系在过去几十年中也发生了明显变化。戈德哈伯建议：“如果你有一个网站，不要收费，因为那只会减少注意力。如果你不知道如何在不收费的情况下负担得起，那么你可能做错了什么。”
It’s useful to reflect on how different the relationships were between economy and attention before the web. In 1997, on the subject of electronic books, Goldhaber wrote, “At present, it’s impractical to distribute books directly over the Internet, though it’s easy to foresee that this won’t be true for long, and physical books will be seen as cumbersome and quaint.” Indeed, over the course of the COVID-19 pandemic, scholars’ dependence on their books and libraries gave way to e-copies and scans thereof, assisting in the writing of the book you hold in your hands (or the e-book you hold in the palm of your hand)—and yet printed book sales exploded. Similarly, the relationship between norms of paywalls has clearly evolved in the intervening decades. Goldhaber advised, “If you have a website, don’t charge for it, because that will just reduce attention. If you can’t figure out how to afford it without charging, you may be doing something wrong.”
然而，二十年后，我们在通过（付费）数字订阅服务听音乐的同时写了这本书，并注意到《纽约时报》基于数字付费墙的业务正在增长。“信息想要免费”，这是 20 世纪 90 年代许多人的想法，但现在信息生产公司依赖的信息实际上很昂贵，需要庞大的基础设施来存储和处理。正如网络上的信息可以以多种方式组织一样，其基础设施也可以通过多种方式支付费用，包括但不限于广告。然而，一些既得利益者长期以来一直声称，实际上存在一种互联网上支付信息的可持续方式：基于对用户活动的监控而投放的广告。这一说法的虚假性并不能否定其历史重要性。
However, twenty years later, we write this book while listening to music via a (paid) digital subscription service and note that The New York Times has a growing business based on a digital paywall. “Information wants to be free,” many people believed in the 1990s, but the information-producing companies now depend on information being, in fact, expensive and requiring massive infrastructure to store and process. Just as information on the web can be organized in many ways, its infrastructure can be paid for in many ways, including, but not limited to, ads. Some vested interests, however, have long claimed that there’s really one sustainable way to pay for information on the internet: ads based on surveillance of the activities of users. The falsity of the claim doesn’t negate its historical importance.
注意力经济与其他经济一样，并不是一只纯粹看不见的手；例如，政府施加版权限制，限制某些信息的复制和分发，从而保持其稀缺性，从而保持其价值。国家在设定权力平衡方面的作用将是下一章的核心。15
The attention economy, like others, is not a purely invisible hand; the government imposes, for example, copyright restrictions which limit the copying and distribution of some information, retaining its scarcity and thereby its value. The role of the state in setting the balance of powers will be central in our next chapter.15
如果信息想要免费，谁来付钱？又由谁来构建它？
If Information Wants to Be Free, Who Is Paying? And Who Will Build It?
信息想要免费，因为信息传播、复制和重组的成本已经变得非常低廉——低到无法计量。信息想要昂贵，因为对于接收者来说，信息的价值是不可估量的。这种矛盾不会消失。
Information wants to be free because it has become so cheap to distribute, copy, and recombine—too cheap to meter. It wants to be expensive because it can be immeasurably valuable to the recipient. That tension will not go away.
—斯图尔特·布兰德16
—Stewart Brand16
戈德哈伯发表预言时，斯坦福大学的 Back-Rub/Google 还只是 NSF 资助的一个研究生项目，旨在将网站作者的劳动转化为网页排名算法。但很快，这个项目就变成了风险投资支持的初创企业谷歌，据说它诞生在帕洛阿尔托的一个车库里。最初的论文描述了 PageRank 算法——这项技术进步使谷歌有别于其他许多当代组织网络的创新者——但并没有提到通过广告将该算法所需的基础设施货币化。人们可以想象许多其他收入模式，比如订阅、联盟费用或赞助链接。广告胜出。17
Goldhaber delivered his prophecies at a time when Back-Rub/Google at Stanford was a mere NSF-backed graduate student project to turn the labor of website authors into a ranking algorithm for web pages. Very soon, though, this project became the venture capital–backed start-up Google, proverbially born in a garage in Palo Alto. The original paper describing the PageRank algorithm—the technical advance differentiating Google from scores of other contemporary innovators organizing the web—included no mention of monetizing the algorithm’s needed infrastructure via advertising. One could have imagined many other income models such as subscriptions, affiliate fees, or sponsored links. Ads won out.17
21 世纪初期出现了一种称为“Web 2.0”的技术规范：所有用户都可以成为发布者，只需向网站提供用户生成内容 (UGC)，网站将托管这些内容，以换取这些用户的劳动和创造力，并越来越多地通过将这些用户的劳动和创造力货币化来换取收入。从 1999 年开始，“Web 2.0”一词在 2004 年由 O'Reilly Media 发起的“Web 2.0 大会”之后变得更加突出。18该公司的创始人 Tim O'Reilly 最初从事技术书籍的编写工作，但在 2000 年“互联网泡沫”破灭后，他将自己的商业模式多样化，包括会议和技术书籍；后来，他创立了自己的风险投资公司。各种各样的网站纷纷涌现以托管 UGC，进一步促进了此类内容的激增。这些网站需要结合设计和算法优化来导航，以及支付服务器空间和带宽的资源（尤其是在视频变得流行的情况下）。用户创作内容的使用日益增多——互联网创作民主化承诺的延续——具有讽刺意味的是，新中介机构的出现不仅组织用户的创造力，也越来越多地试图从中获利。
The early 2000s saw the emergence of a technological norm termed “Web 2.0”: the idea that all users could become publishers by providing user-generated content (UGC) to the sites who would host the content in exchange for, and, increasingly, by monetizing, these users’ labor and creativity. Dating from 1999, the phrase “Web 2.0” gained more prominence after the “Web 2.0 Conference” started in 2004 by O’Reilly Media.18 Tim O’Reilly, the company’s founder, began his career writing technical books, but after the “dot-com crash” of 2000, he diversified his business model to include conferences as well as technical books; later, he founded a venture capital firm of his own. A variety of sites proliferated to host UGC, further encouraging the explosion of such content. These sites required a combination of design and algorithmic optimization to navigate as well as resources to pay for server space and bandwidth (especially as video became prominent). The growing use of user-created content—the continuation of the promise of the democratization of creation on the internet—came ironically with the creation of new intermediaries that organized, but also increasingly sought to profit from, their users’ creativity.
整理互联网上的海量信息没有单一的解决方案。在 Reddit 等网站上，社区被组织成不同主题的动态子组（或“subreddits”），帖子根据用户投票进行算法排序。然而，没有这些设计限制和社区劳动，就有机会通过算法对无结构的帖子提要进行排序。虽然一些网站（如“反社会书签网站” Pinboard）选择订阅模式来赚取收入，但到 2000 年代末，广告已成为主流。广告支持的 UGC 托管网站接受了算法挑战，从数十亿条内容中选择要呈现的内容。这些算法与所有机器学习一样，都是优化算法，要求技术人员做出主观的设计选择：需要优化的功能是什么？越来越多的设计师选择在网站上花费的时间（因此在广告中花费的时间）作为需要优化的功能。这个世界充斥着广告支持、算法优化的 UGC，充斥着我们的日常生活。你正沉浸其中。虽然一些读者无疑是在它陪伴下长大的，但它的普遍性不应与必要性混为一谈。万维网的第一个十年并不是这样的。广告战胜其他支付互联网服务的方式，如今被视为自然而然的，甚至是不可避免的。强大的选民一直在努力让我们这样想。
Organizing the deluge of information on the internet had no one solution. On sites such as Reddit, communities are organized into dynamic subgroups (or “subreddits”) of different topics, with posts algorithmically sorted based on user voting. Without these design constraints and community labor, however, there was an opportunity to sort a structureless feed of posts algorithmically. While some sites, like Pinboard, the “antisocial bookmarking site,” chose a subscription model for revenue, the dominant norm by the end of the 2000s was advertising. Ad-tupported UGC-hosting sites took up the algorithmic challenge of choosing which of billions of pieces of content to present. These algorithms, as with all machine learning, are optimization algorithms, requiring technologists to commit to a subjective design choice: What is the function to be optimized? Increasingly, designers chose time spent on a site—and therefore among ads—as the function to be optimized. This world, with ad-supported, algorithmically optimized UGC, saturates our daily life. You are soaking in it. While some readers no doubt grew up with it, its pervasiveness should not be conflated with necessity. This is not the way things were in the first decade of the World Wide Web. The victory of ads over other ways to pay for internet services is today seen as natural, maybe even inevitable. Powerful constituencies have worked hard to make us think this way.
要理解互联网广告的增长，我们需要看到，主要广告商认为网络远不如电视或印刷媒体负责。这在今天看来是极度违反直觉的，但他们认为自己比起网络广告更了解传统媒体广告的成功。在 20 世纪 80 年代和 90 年代，广告公司根据如何最好地接触受众的计算模型展开竞争，这些模型“成为即将到来的无处不在的数字媒体时代的试验场”。19 《广告时代》 1998 年的一篇文章指出，“缺乏准确的衡量标准和难以追踪投资回报是今年早些时候全国广告主协会进行的一项调查中提到的购买网络媒体的最大障碍。” 20正如历史学家 Joseph Turow 所记录的那样，传统广告商通过要求指标和利用技术进行追踪，帮助创建了一个截然不同的网络。一位曾负责宝洁和戴尔业务的大型广告公司负责人表示：“当这种媒介证明自己是负责任的时候，广告商会很乐意在网上投入更多资金。” 21随着互联网的出现，广告商要求提高责任感，要求获得更多的受众数据，使用可以追踪受众注意力的技术。监控并不是新兴互联网公司嫁接到资本主义上的外来特质；它是主要传统广告商和广告公司需求之间的一种舞蹈，这些广告公司推动技术人员满足这些指标，以便对在线广告的有效性负责，而这种要求越来越多地通过开发用户的详细资料来实现。
To understand the growth of internet advertising, we need to see that major advertisers saw the web as far less accountable than television or print media. It’s highly counterintuitive to us today but they believed they understood the success of legacy media advertisements far better than they did online advertising. In the 1980s and 1990s, advertising firms competed based on their computational models of how best to reach audiences, which “served as a testing ground for the coming age of ubiquitous digital media.”19 An article in 1998 in Advertising Age noted, “Lack of accurate measurement and difficulty tracking return on investment are cited as the biggest barriers to buying online media in a survey conducted by the Association of National Advertisers earlier this year.”20 Legacy advertisers helped create a very different web by demanding metrics and leveraging technologies to allow tracking, as historian Joseph Turow has documented. “Advertisers will be excited to spend more money online,” one major agency head who handled Proctor & Gamble and Dell remarked, “when the medium proves it is accountable.”21 With the coming of the internet, advertisers pushed for more accountability, requiring more data on audiences, using techniques that could track their attention. Surveillance wasn’t a foreign quality grafted onto capitalism by the new internet firms; it was something that emerged as a dance between the demands of the major legacy advertisers and advertising firms that pushed technologists to meet those metrics to provide accountability about the effectiveness of on-line ads, increasingly by developing granular profiles of users.
追踪用户需要网络浏览器包含可进行监视的技术——最臭名昭著的是 cookie 和隐藏的追踪像素。许多负责制定网络标准的技术人员很快意识到隐私的丧失，其中一些人试图改变网络浏览器的标准，以便默认为用户提供更多保护。新兴的互联网广告业进行了积极反击，包括游说浏览器制造商。他们最终赢得了这场战斗，让一亿个 cookie 在你的电脑上绽放（烘焙？）。“在我看来，这是一种极端的反应，”一位高管辩称，“来自一群说……‘我们要说服你网络上存在隐私问题’的人。” 22
Tracking users required web browsers to include technologies that enable surveillance—most notoriously cookies and hidden tracking pixels. The loss of privacy quickly became obvious to many technologists responsible for the standards undergirding the web, some of whom sought to alter the standards for web browsers to afford users more protection by default. The nascent internet advertising industry fought back aggressively, including lobbying browser manufacturers. They ultimately won the battle to let a thousand million cookies bloom (bake?) on your computer. “It seems to me it’s an extreme reaction,” one executive argued, “from a bunch of people who are saying . . . ‘We’re going to convince you there is a privacy problem on the Web.’ ”22
当然，网络上存在隐私问题——这就是重点所在。技术专家和隐私活动家被成功地描绘成反对商业的激进分子，但他们输掉了这场战斗。行业反击强调了网络需要广告这一观点——而广告需要能够跟踪用户。保护隐私的负担再次落在了个人用户身上。联邦贸易委员会在 1998 年的一份报告中解释说，“只需点击计算机屏幕上的一个框，用户就可以轻松做出选择，该框指示用户对所收集信息的使用和/或传播的决定。” 23
Naturally, there was a privacy problem on the web— that was the entire point. Successfully depicted as radicals against commerce, the technologists and privacy activists lost this battle. Industry pushback emphasized the idea that the web required advertising—and advertising needed to be able to track users. Once again, the burden of protecting privacy was left to individual users. In a 1998 report the Federal Trade Commission explained, “choice easily can be exercised by simply clicking a box on the computer screen that indicates a user’s decision with respect to the use and/or dissemination of the information being collected.”23
最臭名昭著的是，1995 年成立的 Doubleclick 将销售广告与收集数百万用户的数据联系起来。DoubleClick 总裁兼首席执行官凯文奥康纳解释说：“定向广告的最大悖论是，你越是进行微定位，你就必须拥有更大的覆盖范围”——你越想让广告有针对性，你就越需要收集每个用户的大量数据。24尽管隐私权倡导者和政府监管机构在2000年左右对 DoubleClick 提出了挑战，但该公司出色地经受住了对其行为的有限限制。25可以肯定的是，该公司提供了选择退出的选项，但正如马修·克雷恩 (Matthew Crain) 指出的那样，默认情况下，少数“选择退出的选项只是监控海洋中的一滴水”。26到20世纪 90 年代末，网络广告已成为常态，用户浏览时缺乏隐私也已成为常态。在接下来的十年里，默认采用广告模式也成为大量“颠覆者”的共同规范，Facebook 和许多其他现已成功的公司都选择广告模式作为他们风险投资后的命脉。
Most infamously, Doubleclick, founded in 1995, connected selling ads to collecting data on millions of users. DoubleClick president and CEO Kevin O’Conner explained, “The great paradox with targeting ads is that the more you are micro-targeting, the more reach you have to have”—the more you want to focus ads, the more you need to collect large amounts of data on each user.24 While privacy advocates and government regulators alike challenged DoubleClick around 2000, the firm weathered the limited restrictions on its practices strikingly well.25 The company, to be sure, made opting out possible, but, as Matthew Crain notes, the small number of “opt-outs were a drop in an ocean of surveillance” by default.26 By the end of the 1990s, advertising on the web had become the norm, as had the lack of user privacy when browsing. Over the next decade, defaulting to the ad model would become the common norm among a flood of “disruptors” as well, with Facebook and many other now-successful companies choosing the ad model as their post–venture capital lifeblood.
谷歌的广告最初侧重于搜索词和上下文，而不是用户监控。它的整个商业模式很快就经历了一场巨大的转变。2005 年，谷歌收购了一家旨在提供基础设施以将广告和用户网络货币化的初创公司，这家初创公司本身已经被私募股权公司收购——Doubleclick。2009 年 3 月 11 日，谷歌宣布监控广告将成为其未来：基于兴趣的广告“将根据您访问的网站类型和您查看的页面，将兴趣类别（例如体育、园艺、汽车、宠物）与您的浏览器相关联。然后，我们可能会使用这些兴趣类别向您展示更相关的文字和展示广告。” 27
Advertising on Google originally focused on search terms and context, not surveillance of users. Its entire business model soon underwent a dramatic transformation. In 2005, Google purchased a company founded to provide infrastructure to monetize the network of advertisements and users, a start-up which itself had already been acquired by a private equity company—Doubleclick. And on March 11, 2009, Google announced that surveillance advertising was to be its future: interest-based ads “will associate categories of interest—say sports, gardening, cars, pets—with your browser, based on the types of sites you visit and the pages you view. We may then use those interest categories to show you more relevant text and display ads.”27
到 2000 年代中期，很明显，拥有允许任何人在其他平台上买卖广告的技术将为任何能够占据主导地位的中介带来极其丰厚的利润。到那时，广告模式已被内容生产-消费交换中的相关各方所接受：创始人、投资者和用户（尽管不情愿）。这对公司创始人来说已成常态；例如，谷歌的拉里·佩奇和谢尔盖·布林，以及 Facebook 的马克·扎克伯格，不得不默许他们主要通过销售广告来赚钱。不久之后，他们就成为网络广告的主导平台。投资者必须首先发现这一点是合理的，然后才是规范的。用户必须同意，即使是无意识的。如果用户集体拒绝一个广告支持的网站，因为它上面有广告，那么这个网站就完了；如果用户耸耸肩，接受网站展示广告是正常的，尽管这对用户体验有有害和干扰性的影响，那么这是一种可行的商业模式。同样，如果内容生产者——所有发布、发推文和上传他们的希望、表情包和恐惧的人——同意免费这样做，那么 UGC 将继续作为一种产品-技术-商业模式蓬勃发展。我们每个人都会做出这些选择，而这种商业模式的延续需要我们的公共规范和市场不断发展，跟上私营公司技术架构的变化。
By the middle of the 2000s, it was clear that having the technology to allow anyone to buy and sell ads on other platforms was going to be extremely profitable for whomever could dominate as intermediaries. By this time, the ad model was accepted by the relevant sides in the content producing-consuming exchange: the founders, the investors, and the users (albeit begrudgingly). It became normal for the founders of the companies; e.g., Google’s Larry Page and Sergey Brin, and Facebook’s Mark Zuckerberg, had to acquiesce that they were going to make money primarily by selling ads. Before long, they became the dominant platforms for web advertising. Investors had to find this first plausible and then normative. And users had to consent, even if unconsciously. If users collectively reject an ad-supported website because it has ads on it, that’s the end of the website; if users instead shrug and accept that it’s normal for sites to show ads despite the deleterious and interruptive effect on user experience, then it’s a plausible business model. Similarly, if content producers—all those who post and tweet and upload their hopes, memes, and fears—agree to do so for free, then UGC continues to thrive as a product-technology-business model. Each of us makes these choices, and the continuation of such business models requires our public norms and markets to evolve, keeping up with private companies’ changes in technical architecture.
数据和广告
DATA AND ADS
让我们再来看一下上面提到的广告定位技术挑战：如何实现这一目标？创建 DoubleClick 这样的广告交易平台需要什么？它需要软件来显示广告，还需要投资机器学习来确定向谁展示广告以及如何定价。到 20 世纪 90 年代末，宝洁等大型广告商担心互联网广告公司承诺很多但尚未兑现。广告商担心他们对传统媒体受众的了解多于对网络用户的了解。“为了证明其价值，”传播学者 Matthew Crain 解释说，“互联网广告行业需要提高其针对特定消费者群体的能力，并证明在线广告可以推动消费者的消费行为。” 28为了证明他们可以提供更多的投资回报，这些公司声称他们需要更多的用户数据以及更好的机器学习。
Let’s double-click on the technological challenge of ad targeting we mentioned above: How would you get this done? What is necessary to create an ad exchange like DoubleClick? It requires software to display ads, but also an investment in machine learning to determine to whom to show ads and how to price them. By the late 1990s, large advertisers like Proctor & Gamble worried that internet advertising firms were promising much but weren’t yet delivering. Advertisers worried that they knew more about the audiences of traditional media than they did about web users. “To prove its worth,” communications scholar Matthew Crain explains, “the internet advertising industry needed to improve its capacity to target specific groups of consumers and demonstrate that online ads could move the needle of consumer behavior.”28 To demonstrate they could provide more return on investment, these firms claimed they needed more data on users—as well as better machine learning.
网站通过向不同的人展示不同的广告进行实验，并了解哪些广告将带来营销人员想要的结果——通常是点击或购买。进行这些实验并将其付诸实践需要大量的数据、高性能算法和强大的软件工程。当时撰写这篇文章的人是杰夫·哈默巴赫 (Jeff Hammerbacher)，他曾短暂 (2006-2008) 在 Facebook 工作。哈默巴赫与 Facebook 创始人马克·扎克伯格 (Mark Zuckerberg) 一起就读于哈佛大学，最初是 2004 届。毕业后，哈默巴赫去了贝尔斯登 (Bear Stearns)，对金融深感厌倦，然后转到 Facebook 工作了两年，帮助当时的老板亚当·德安杰洛 (Adam D'Angelo)。德安杰洛被誉为 Facebook“增长团队”的创始人，他后来将其描述为“一支工程师团队，他们以各种方式改变产品，使其更具病毒性并吸引更多用户注册。” 29正如第 10 章所讨论的，哈默巴赫在 2011 年的一次采访中描述了当时的情形，“我们这一代最优秀的人才正在思考如何让人们点击广告。这太糟糕了。” 30 Facebook 和谷歌后来成为当今最大的两家公司，共同主宰着数字广告。
Sites experiment by showing different ads to different people and learning which ads are going to drive outcomes desired by marketers—usually a click or purchase. Undertaking these experiments and putting them into practice required abundant data, performant algorithms, and robust software engineering. One person who wrote about this at the moment was Jeff Hammerbacher, briefly (2006–2008) of Facebook. Hammerbacher attended Harvard with Facebook founder Mark Zuckerberg, originally in his class of 2004. After graduating Hammerbacher went to Bear Stearns, became deeply bored with finance, and then moved to Face-book for two years, helping his then-boss Adam D’Angelo. D’Angelo is credited with creating Facebook’s “growth team,” which he later described as “a team of engineers who changed the product in various ways to make it more viral and get more users to sign up.”29 As discussed in chapter 10, Hammerbacher described this time in 2011 in an interview, “best minds of my generation are thinking about how to make people click on ads. That sucks.”30 Facebook and Google would go on to become two of the largest companies today, collectively dominating digital advertising.
“在广告上我浪费了一半的钱。问题是我不知道是哪一半。”这句俏皮话通常被认为是广告商约翰·沃纳梅克说的，其文化影响力可以从纽约市以他的名字命名的街道来衡量，后来这条街道成为美国在线和尼尔森的总部，也是 Facebook 纽约办事处的所在地。然而，在数字领域，营销人员努力想知道他们浪费了哪一半：自 20 世纪 90 年代末以来，点击次数很容易记录，用户也很容易跟踪，软件可以部署使用 RA Fisher 在 1925 年推崇的相同科学：随机对照试验，以了解几种“治疗方法”（这里指广告）中的哪一种优化参与度。事实上，只需多花一点算法上的功夫，营销人员不仅可以了解什么是最好的（描述性分析），还可以优先投放参与度更高的广告（规范性分析）。现在，在网络编程中，利用可以追溯到 1933 年的数学方法，可以很容易地做到这一点。31
“In advertising I’m wasting half my money. The problem is I don’t know which half.” Often attributed to the advertiser John Wanamaker, this quip’s cultural impact can be measured by the street named after him in New York City, later the headquarters of AOL and Nielsen, and the site of Facebook’s NYC office. In the digital domain, however, marketers fought hard to know which half they were wasting: since the late 1990s, clicks have been easy to record and users to track, and software may be deployed to use the same science extolled by R. A. Fisher in 1925: the randomized controlled trial, to learn which of several “treatments” (here, advertisements) optimizes engagement. In fact, only a bit more algorithmic effort allows marketers not only to learn what is best (a descriptive analysis) but preferentially to serve the ad with greater engagement (a prescriptive analysis). Simple methods for doing so are now commonplace in web programming, exploiting mathematical methods dating as far back as 1933.31
“让人们点击广告”需要什么？首先，需要大量丰富多样的新鲜内容，而这些内容目前主要由用户生成内容提供。在 Facebook 的 Newsfeed、TikTok 或 YouTube 的推荐视频等数字产品中，选择分享数十亿个来源中的哪一个是这些算法的主要输出。这还需要参与度的历史数据——类似的先前用户对先前内容的反应（点击、分享、喜欢等）。回想一下，自 1945 年技术人员试图摆脱打孔卡以来，在存储和使用信息方面取得了多大的进步。个性化这种优化需要有关用户的信息；虽然设备类型或地理位置等“粗略”信息可能对优化有用，但几十年来广告中人们所理解的人口统计信息不仅对优化有用，而且对营销人员也有吸引力。也就是说，拥有技艺和领域直觉的营销人员希望在特定的“客户群体”上花钱，比如“NASCAR 妈妈”或“足球爸爸”（用《纽约时报》政治营销人员的例子来说），因此对于广告平台来说，用这种语言描述用户并预测用户是否属于这些群体是很有用的。
What is needed to “make people click on ads”? Above all, a tremendous abundance of varied and fresh content, provided now largely by user-generated content. Choosing which of these billions of sources to share is the main output of these algorithms in digital products like Face-book’s Newsfeed or TikTok or YouTube’s recommended videos. This also requires historical data on engagement— the response of similar prior users to prior content (clicks, shares, likes, etc.). Recall what an advance in storing and making information usable has been achieved since 1945, when technologists were attempting to move off punch cards. Personalizing this optimization requires information about users; while “coarse” information like the device type or the geographic location can be useful to optimization, demographic information as had been understood in advertising for decades is not only useful for optimization but saleable to marketers. That is, marketers possessing a craft and domain intuition hope to spend on a particular “customer segment,” such as “NASCAR moms’’ or “soccer dads,” to use examples from a political marketer in The New York Times, and so it’s useful for advertising platforms to describe users in this language and to model predictively whether users fall into these groups.
广告与现实
Advertising and Reality
广告，尤其是其最近的机器学习优化形式，对于我们构建的方式意味着什么？我们能从我们感知的世界里看出现实吗？我们掌握在手中的真相主要来源是由监控广告模式资助和优化的，这意味着什么？我们不必接受对广告或宣传权力的肤浅看法，就能认识到掌握文化中介和选择系统的必要性。
What does advertising and, in particular, its recent machine learning optimized form mean for the way that we all construct reality from the world we perceive? What does it mean that our primary source of truth, delivered to us in the palms of our hands, is funded by, and optimized for, the surveillance ad model? We don’t have to accept a facile view of the power of advertising or of propaganda to recognize the need to grasp the systems of cultural mediation and selection at work.
为了更好地理解这一点，我们首先来讨论一下优化对于信息和说服的意义。这个主题将设计、数据和激励结合在一起；最终，它还涉及讨论谁在进行优化以及为什么进行优化，这将使我们了解业务目标如何推动设计选择，进而塑造我们对现实的看法。
To understand this better, first let’s discuss what optimization means for information and persuasion. This topic brings together design, data, and incentives; ultimately, it also involves discussion of who is doing the optimizing and why, which will bring us to understand how the business goals drive design choices, which in turn shape our perception of reality.
就像为优化手写识别或预测 Netflix 的电影评论而引入的常见任务框架一样，将广告简化为简单的指标，使数据科学家和产品开发人员能够将机器学习的武器库转向优化商定的广告指标。这种动态的一个例子是常见的数字广告指标 CPM 或“每千次展示费用”。购买像素的营销人员选择将他们的广告投放在不同平台上，部分原因是 CPM 较低；出售这些像素的广告商则寻求最大化总展示次数。（需要明确的是，并非所有数字广告交易都是如此定价和谈判的，但这种框架是对程序化广告和广告交易网络的有用简化。）
Like the common task frameworks introduced for optimizing handwriting recognition, or predicting movie reviews at Netflix, the reduction of advertising to simple metrics has allowed data scientists and product developers to turn the armamentarium of machine learning toward optimizing agreed-upon advertising metrics. An example of this dynamic is illustrated by the common digital advertising metric CPM or “cost per thousand impressions.” Marketers, who buy pixels, choose to place their advertisements on different platforms based in part on low CPMs; advertisers, who sell these pixels, instead seek to maximize the number of total impressions. (To be clear, not all digital advertising deals are so priced and negotiated, but this framing is a useful simplification of programmatic advertising and ad exchange networks.)
优化展示次数意味着让用户查看尽可能多的内容，并尽可能多地返回使用产品。与任何其他常见任务框架一样，这会改变人们看重的东西——不是用户的幸福感，也不是信息的真实性，而是用户总数展示次数。同样，如果在用户使用产品时以规律的节奏投放广告，这意味着优化“北极星”指标，通常称为“参与度”。对于信息平台而言，参与度越高，收入也就越多，无论向用户展示的信息性质如何。
Optimizing impressions means getting users to view as much content as possible, and to return to use the product as much as possible. Like any other common task framework, this changes what is valued—not the happiness of the user, or the veracity of the information, simply the total number of impressions. Similarly, if ad impressions are delivered with a regular cadence as the user uses the product, this means optimizing a “North Star” metric for use, often called “engagement.” For information platforms, then, more engagement means more money, irrespective of the nature of the information shown to the user.
尤其是随着从台式电脑转向移动端，信息平台的兴起要求视觉设计选择能够在小空间（通常是手掌）内实现最大信息流。这包括剥离上下文线索，例如在聚合内容提要中提供信息来源的详细信息，以及重要的是混合各种类型的内容：新闻、娱乐、朋友的帖子、陌生人的帖子，当然还有广告。Facebook 于 2006 年推出的新闻提要就是这种聚合的典型例子。内容可以来自网络上数十亿用户生成的帖子。这是波兹曼所警告的状态的顶峰：“新闻”，在善意尝试向社区中的每个人提供事实、有用的信息方面，不再与赞助内容、说服性内容、虚假或讽刺内容和娱乐分开。32
Particularly with the shift to mobile from desktop computing, the rise of information platforms has required visual design choices which facilitate maximum information flow in a small space—often the palm of one’s hand. This includes stripping away contextual cues, such as the details of the source of the information when served in an aggregating feed of content as well as, importantly, mixing types of content: news, entertainment, posts from friends, posts from strangers, and, of course, advertisements. Facebook’s News Feed, which debuted in 2006, exemplifies such aggregation. Content could be sourced from billions of user-generated posts from around the web. This is the culmination of the state Postman warned about: “news,” in the sense of good faith attempts to provide factual, useful information to everyone in a community, no longer is separated from sponsored content, persuasive content, false or satirical content, and entertainment.32
这些设计决策可以按照 Fisher 在 1925 年倡导的方式进行优化，使用随机对照试验来推动北极星指标。业内称之为“A/B 测试”。参与度的真正提升来自于通过算法选择向哪些用户展示哪些内容。信息流的设计打破了广播或电视等媒体的节目限制，将任何形式的引人入胜的内容与任何其他内容和赞助内容相邻呈现。传统媒体面临的另一个限制是“一对多”限制，这意味着每个人都看到相同的内容。信息平台优化不受此限制的束缚，为每个人提供最具吸引力、独特、定制的现实。
These design decisions can be optimized, in the same way Fisher advocated in 1925, using randomized controlled trials to drive a North Star metric. Industry terms these “A/B tests.” The real gain in engagement comes from algorithmically choosing which content to show which users. The design of the feed breaks the programming constraints of media such as radio or television, by presenting any form of engaging content adjacent to any other and to sponsored content. An additional constraint faced by traditional media was the “one to many” constraint, meaning that everyone saw the same content. Information platforms optimize unfettered by this constraint, providing optimally engaging, distinct, bespoke realities to each individual.
多年来，技术专家和非技术专家越来越多地注意到，哪些类型的内容最吸引人。正如波兹曼所警告的那样，最吸引人的内容不一定是最真实或最有用的。法学教授詹姆斯·格里梅尔曼在他的文章《平台即信息》（2018 年）中引用了马歇尔·麦克卢汉的著名格言“媒介即信息”，研究了这一动态。平台“一直在仔细观察，看看哪些内容在吸引注意力方面胜过竞争对手”，格里梅尔曼解释说。“这些平台优先考虑和推广最有可能吸引用户的内容。”虽然 Udny Yule 只有观察数据来为政策提供信息，但网站可以——而且一直在做——进行干预并记录由此产生的效果，而不仅仅是观察活动。当你使用数字产品时，你一直在经历一场涉及你的实验。就用户生成内容 (UGC) 而言，向 feed 提供哪些内容的选择非常多，甚至达到数十亿。如果你是一名专业的 YouTuber（这样的人很多），你会仔细关注你的数据。想象一下，你发现自己关于某个新话题的视频吸引了很多人的关注。这是一种激励，促使你去制作更多有关该话题的视频。其结果是，几乎所有用户群体的品味都会得到提升。正如格里梅尔曼所观察到的，“推荐引擎可能只会推荐车祸视频，因为你看了这一段，所以这里还有另一段。你可能有兴趣看那段。” 33他指出，现在会自动生成内容以满足新产生的兴趣。即使这些平台不能简单地销售商品或影响我们的政治观点，但以参与度为核心的平台也会以不可预测的方式改变着我们的信息世界。
Increasingly, technologists and non-t echnologists have noted patterns over the years in what types of content are the most engaging. As Postman warned, the most engaging content is not necessarily the most factual or useful. Alluding to Marshall McLuhan’s famous maxim, “the medium is the message,” law professor James Grimmelmann investigated this dynamic in his essay “The Platform Is the Message” (2018). Platforms “are carefully and constantly watching to see which content beats out its rivals in drawing attention,” Grimmelmann explains. “These platforms prioritize and promote the content most likely to grab users by the lapels.” While Udny Yule had only observational data to inform policy, websites can—and constantly do—perform interventions and record the resulting effects, rather than merely observing activities. When you use a digital product, you are experiencing an experiment involving you all the time. In the case of user-generated content (UGC), the choices of which content to supply to a feed are immense, well into the billions. When you’re a professional YouTuber (and they are legion), you’re carefully watching your numbers. Imagine that you find that your videos on a new topic draw a lot of attention. It is an incentive to go out there and produce more on that topic. The result is to amplify nearly any taste of any user segment. As Grimmelmann observes, “Recommendation engines may only supply car crashes in the sense of suggesting that, since you looked at this one, here’s another one. You may be interested in watching that.”33 And—he notes—content will now be automatically produced to cater to newly created interests. Even if they don’t work to vend goods or sway our politics in a simple manner, platforms organized around engagement alter our informational world in unpredictable ways.
难道不能用更多的人工智能来治愈这个人工智能吗？
CAN’T ONE HEAL THIS AI WITH MORE AI?
为什么我们不能使用监督学习（在没有治疗的情况下预测结果）或强化学习（选择最佳治疗以最大化结果）来解决“麻烦内容”（例如假新闻、错误/虚假信息、仇恨言论和辱骂）问题？挑战包括公司消除此类引人入胜内容的动机（再次引用 Grimmelmann 的话：“抱怨也无济于事。仇恨点击仍然是点击”），以及缺乏明确的标签。即使有成千上万的版主（例如 Facebook 的情况），或者即使有大量众包“幽灵劳动”，不同的读者也经常会对讽刺、模仿和善意或恶意在线论点的性质产生分歧。例如，Facebook 的“监督委员会”经常被要求对特别令人不安的内容作出裁决。作为一个“人工智能”问题，我们受到缺乏甚至实际智能的明确判断的限制。
Why can we not use supervised learning, which predicts an outcome in the absence of treatment, or reinforcement learning, which chooses the best treatment to maximize an outcome, to fix the “troublesome content” (e.g., fake news, mis/disinformation, hate speech, and abuse) problem? Challenges include the incentives of the companies to eliminate such engaging content (again, Grimmelmann: “Complaining about it doesn’t help, either. Hate clicks are still clicks”), but also the lack of clear labels. Even with tens of thousands of moderators, as in the case of Facebook, or even with abundant crowdsourced “ghost labor,” different readers will often disagree as to the nature of satire, parody, and good- or bad-faith arguments online. Facebook’s “Oversight Board,” for example, is often called upon to adjudicate decisions over particularly troubling content. As an “AI” problem, we are limited by the lack of clear judgment of even actual intelligences.
内容审核的另一个难题是：即使平台公司确实采取立场，禁止或只是降低令人不安的内容的显示频率，这种行为也会招致偏见或审查的指责，而这反过来又会引起人们对他们不想传播的内容的关注，即扩大这些内容。将内容标记为有问题的设计选择也有缺点，包括引起人们对此类标签的关注以及“适得其反效应”，这种现象有时会在实验中观察到，在实验中，纠正信息会导致人们对矛盾的说法更加信任。
An additional conundrum in content moderation: even when platform companies do take a stand, banning or simply down-weighting the frequency at which troubling content is shown, such actions draw accusations of bias or censorship which in turn can draw attention to—i.e., amplify—the content they sought not to distribute. Design choices which label content as problematic have drawbacks as well, including both the attention drawn to such labels as well as “backfire effect,” a phenomenon sometimes observed in experiments in which correcting information results in increased faith in the contradicted claim.
民主化说服，从营销到政治
DEMOCRATIZED PERSUASION, FROM MARKETING TO POLITICS
Zeynep Tufekci 称之为复合“说服架构”，包括机器学习算法以及实例化算法的产品。这些架构统计性能和易用性都已优化到任何人都能做到的程度。例如，Facebook 的“相似受众”功能允许营销人员要求 Facebook 找到与执行了特定操作的人相似的其他人。营销人员因此可以将内容定位到在人口统计学或行为上与执行了特定操作（例如点击特别挑逗性的链接）的先前用户相似的新用户。这无需任何关于市场研究或用户心理学的特别智慧或直觉即可完成；对于个人购买者来说，只需花费少量资金，但对于 Facebook 或 Google 来说，这笔钱总计巨大。
The composite “persuasion architecture,” as Zeynep Tufekci terms it, includes the machine learning algorithm as well as the product that instantiates the algorithms. These architectures have become so optimized in statistical performance as well as ease of use that anybody can do it. As one example, Facebook’s “lookalike audiences” allows marketers to ask Facebook to find other people who look like the people who performed a certain action. Marketers can thus target content to new users who resemble demographically or behaviorally prior users who performed an action, like clicking on a particularly provocative link. This can be done without any particular savvy or intuition about market research or user psychology; just a small amount of money for the individual purchaser, though vast in aggregate for Facebook or Google.
早在 20 世纪 20 年代，伯内斯就意识到，说服法则在营销和政治领域同样适用。在伯内斯磨练烟草信息之前，他正在塑造公众对政治的看法。1924 年，他安排一群受欢迎的名人与“几乎不善言辞”的总统卡尔文·柯立芝一起出现，以提升他的形象。今天，有抱负的政治家从与营销人员相同的统计方法中受益，利用数字工具进行市场调查，以策划活动和信息，并将最佳信息传递给最佳用户。伯内斯认为广告和政治之间没有界限；事实上，他认为这种同意工程对民主来说是一件好事。“伯内斯认为这是任何民主不可避免的一部分，”泽伊内普·图费克奇说。 “他和杜威、柏拉图和李普曼一样，相信强者比大众具有结构性优势……他敦促善意的、技术和经验上可行的政治家通过操纵和同意工程技术成为‘哲学王’。” 34
Bernays realized, even in the 1920s, that the laws of persuasion apply equally in marketing or in politics. Before Bernays was honing the message for tobacco, he was shaping public perception in politics. In 1924 he arranged for a group of popular celebrities to appear with the “practically inarticulate” president Calvin Coolidge to improve his image. Today, aspiring politicians benefit from the same statistical methods as marketers, leveraging digital tools for market surveys to craft campaigns and messages as well as to target the optimal message to the optimal user. Bernays saw no line between advertising and politics; indeed, he saw this engineering of consent as a good thing for democracy. “Bernays saw this as an unavoidable part of any democracy,” according to Zeynep Tufekci. “He believed, like Dewey, Plato and Lippmann had, that the powerful had a structural advantage over the masses. . . . He urged well–meaning, technologically and empirically enabled politicians to become ‘philosopher–kings’ through techniques of manipulation and consent engineering.”34
虽然这些技巧可以而且必须用于善事，但伯内斯很清楚，它们“可以被颠覆；煽动家可以利用这些技巧来达到反民主的目的，其成功程度与那些利用这些技巧来达到社会期望的目的的人一样高。目的。”伯内斯认为，一个追求良好目标的领导人必须“把精力投入到掌握同意工程的操作诀窍上，并为了公众利益智胜对手。” 35尽管伯内斯对“宣传”一词的坦诚使用在冷战期间已失宠，但随着数据和算法越来越多地用于传递和优化说服性信息，“工程同意”的有效性却急剧上升。不久之后，在线数字广告的完善技术就开始为政治信息传递提供信息。
While these techniques can—and must be—used for good, Bernays was clear, they “can be subverted; demagogues can utilize the techniques for antidemocratic purposes with as much success as those who employ them for socially desirable ends.” A leader seeking good objectives, Bernays argued, must “apply his energies to mastering the operational know–how of consent engineering, and to out–maneuvering his opponents in the public interest.”35 Even as Bernays’s bracingly honest use of the term “propaganda” fell out of favor during the Cold War, the effectiveness of “engineering consent” grew dramatically with the increasing use of data and algorithms to inform and optimize persuasive messaging. It was not long before the techniques perfected for digital advertising online came to inform political messaging.
2007 年，时任埃森哲技术实验室员工的雷伊德·加尼 (Rayid Ghani) 描述了一种“个性化促销计划系统”。他盛赞数据将如何实现全新的个性化定位。“除了使用报纸、店内展示和端盖来突出他们的产品和进行促销之外，零售商还可以使用个人消费者模型，以截然不同的方式影响个人。” 36通过在细微层面上了解客户，可以实现业务目标。这项技术允许每家公司将每个客户视为单独的个体，而不仅仅是统计类别的代表。这种定位正是市场营销和政治运动的核心。37加尼在 2008 年大选中担任奥巴马的首席科学家，部分原因是利用这些数据将选民视为客户细分。
In 2007, Rayid Ghani, then an employee at Accenture Technology Labs, described an “Individualized Promotion Planning system.” He celebrated how data would allow radically new forms of individualized targeting. “In addition to using newspapers, in-store displays, and end caps to highlight their products and run promotions, retailers can influence individuals in a vastly different way using individual consumer models.”36 Business goals can be met by understanding customers at a granular level. This technology allows every company to target every customer as a separate person, not just a representative of a statistical category. Just such targeting is at the heart of marketing and political campaigns.37 Ghani served as Obama’s chief scientist in the 2008 election, in part using these data-empowered views of the electorate as customer segments.
2012 年奥巴马连任后不久，伊桑·罗德在《纽约时报》发表专栏文章，赞扬奥巴马竞选数据战略的核心是个人。他说：“竞选活动……正在朝着将每位选民视为独立个体的方向发展。” 38在过去几年中，人们担心这项技术过于具有说服力，因此对这种乐观的说服力进行了更为悲观的解读。无限的粒度、关于个人的深层背景信息以及个性化很多人担心，为参与而优化的说服架构取得了太大的成功，以致民主无法健康运作。
In an op-ed in The New York Times soon after the reelection of Obama in 2012, Ethan Roeder concluded with a celebration of the individual at the heart of the Obama campaign’s data strategy. “Campaigns are . . . moving toward treating each voter as a separate person.”38 This rosy picture of persuasion is interpreted more darkly after the concerns of the past few years in which this technology has been characterized as too persuasive. Unbounded granularity, deep contextual information about individuals, and personalized persuasion architectures which optimized for engagement had created, many feared, too great a success for democracy to function healthily.
这种被精心设计的公众形象让一些人充满希望，而另一些人则担心当权者侵犯了知情同意权。这些工具很容易被滥用，我们现在知道它们已经被滥用了。图费克奇在 2018 年警告说：“为了通过广告精准定位个人，当今的平台会大规模监视其用户；然后他们使用参与度提升算法让人们尽可能长时间地留在网站上。现在很明显，这个系统很容易被专制、操纵和歧视性地使用”，她举了许多例子。39但除非你根据他们的身份或他们所做的事情来区分他们，否则你无法精准定位某人。而这需要你拥有丰富的数据和机器学习。虽然许多人认为我们的自由市场规范不会受到强大公司提供高度个性化广告的威胁，但当这些能力赋予国家权力时，我们对权力的担忧就会加剧。然而，该算法对国家和公司都有效。40马修·萨尔加尼克（Matthew Salganik）警告我们：“这些能力的变化速度比我们的规范、规则和法律更快。” 41我们还应该补充一点：它们的变化速度很可能比我们用来理解社会和经济现实与概念世界（以平台为媒介，我们通过平台体验和行动）之间关系的分析工具的变化速度更快。
This vision of an engineered public fills some with hope, and others with concern that informed consent has been infringed by those in power. Such tools could easily be put to ill use, and we now know they have. Tufekci warned in 2018, “To microtarget individuals with ads, today’s platforms massively surveil their users; then they use engagement-juicing algorithms to keep people onsite as long as possible. By now it’s clear that this system lends itself to authoritarian, manipulative, and discriminatory uses,” of which she gave numerous examples.39 But you can’t microtarget somebody unless you differentiate them based on who they are or what they’ve done. And that requires you to have copious data and machine learning. While many feel that our free-market norms are not threatened by powerful companies delivering deeply personalized advertising, our concerns over power are heightened when these abilities empower the state. The algorithm, though, works for states and corporations alike.40 Matthew Salganik warns us, “These capabilities are changing faster than our norms, rules, and laws.”41 We should add: they are likely changing faster than our analytical tools for understanding the relationship between our social and economic realities and the conceptual worlds, mediated in platforms, through which we experience and act in them.
肯尼亚 2017 年大选、英国脱欧和 2016 年美国大选极大地放大了人们对算法操纵的担忧，并使之流行起来。引发担忧的焦点是剑桥分析公司；时任首席执行官亚历山大·尼克斯 (Alexander Nix) 在 2017 年分享了一种世界观，与伯内斯的观点相呼应，并进行了更新：
Kenya’s 2017 election, Brexit, and the 2016 US election greatly amplified—and popularized—concerns about algorithmic manipulation. A flashpoint of concern was the firm Cambridge Analytica; then-CEO Alexander Nix shared a worldview in 2017 that echoed and updated that of Bernays:
毫无疑问，营销和广告界领先于政治营销和政治传播领域。我肯定会说，我们所做的一些事情让我感到非常自豪，这些事情是创新的。有些事情是最佳的数字广告实践，最佳的传播实践，我们从商业世界吸取并运用到政治领域。42
There’s no question that the marketing and advertising world is ahead of the political marketing and political communications world. And there are some things that I would definitely [say] I’m very proud of that we’re doing which are innovative. And there are some things which is best practice digital advertising, best practice communications which we’re taking from the commercial world and are bringing into politics.42
研究人员尚未就此次操纵尝试的最终效果达成共识。
Researchers have reached no consensus on the ultimate effects of this attempted manipulation.
担心广告技术和说服架构的影响并不需要我们相信广告商和技术江湖骗子对其广告效果的说法。Tim Hwang 和 Cory Doctorow 精辟地强调了定向广告的深层限制和欺骗。*虽然广告技术，无论是用于商业还是政治，肯定不会像那些兜售它的人所暗示的那样发挥作用，但它极大地改变了我们的媒体格局，并将数字广告商的格局整合为近乎双头垄断的格局（Facebook 和 Google），其影响难以预测。Facebook 和 Google 不需要广告像承诺的那样发挥作用——他们需要广告商相信它们有效。也许这是一场骗局，但它正越来越主宰着我们的信息格局，无论是好是坏。
Worrying about the effects of the adtech and persuasion architectures does not require us to believe the claims of advertisers and technical snake-oil salespeople about the effectiveness of their ads. Tim Hwang and Cory Doctorow have brilliantly stressed the deep limits and deceptions around targeted advertising.* While adtech, either for commerce or politics, surely doesn’t work in the ways those hawking it suggest, it has dramatically transformed our media landscape and consolidated a landscape of digital advertisers into a near duopoly (Facebook and Google), with unpredictable effects. Facebook and Google don’t need the ads to work as promised—they need advertisers to believe they work. Perhaps it’s a shell game, but it’s one that dominates our informational landscape ever more, for better and worse.
接下来，我们讨论当广告模式（收入的规范）遇到风险投资模式（加速市场创新的过程）时会发生什么，我们将看到，这个过程比规范和法律的适应速度更快。
We next discuss what happens when the advertising model—a norm for revenue—meets the venture capital model—a process for accelerating market innovation which, we will see, moves faster than norms and laws adapt.
快速行动：风险投资
Moving Fast: Venture Capital
风险投资甚至不是一项本垒打生意。它是一项大满贯生意。
Venture capital is not even a home run business. It’s a grand slam business.
—风险投资公司 Benchmark 43的普通合伙人 Bill Gurley
—Bill Gurley, general partner at the venture capital firm Benchmark43
福特汽车公司于 1916 年推出了量产汽车，但花了数年时间才制定出将汽车融入社会的规范，并花了数十年时间制定消费者保护法规，例如安全带法。与软件和信息技术今天迅速颠覆我们的规范的方式相比，这种创新的时间表似乎很古怪，因为技术、市场、规范和法律需要数十年才能达到平衡。风险投资 (VC) 有助于加速这种颠覆，因为大规模增长可以先于收入，而对于面向消费者的公司来说，规范（尤其是监管）会减缓新产品的采用。
The Ford Automotive Company introduced mass-produced cars in 1916, but it took years to develop norms around integrating cars into society and decades for regulation about consumer protection, such as seat belt laws. This timescale for innovation, in which technology, markets, norms, and laws have decades to equilibrate, seems quaint compared to the way software and information technology rapidly upends our norms today. Venture capital (VC) has helped to expedite this disruption in that massive growth can precede revenue and, in the case of consumer-facing companies, norms (and, even more so, regulation) that otherwise slow adoption of a new product.
风险投资一直伴随着我们吗？尽管投资一直存在，但许多人指出，第二次世界大战是风险投资模式诞生的时期之一。44例如， 1946年，后来成为哈佛商学院教授的乔治·多里奥特 (George Doriot) 是二战期间负责后勤的军需官。战后，多里奥特创立了 ARDC，这是一家上市公司，投资于长期研究开发，包括许多新兴计算行业的开发。随后几十年来，微处理器和个人计算机在很大程度上都是通过风险投资获得资金的；鉴于个人计算机对社会和经济的积极变化，风险投资受到了广泛的赞扬。在这条值得庆祝的时间线上，缺少了贯穿前几章的数十年的军方对计算的资助，以及通过 SBIR 计划对创新型小型企业的大规模支持，政府实际上充当了风险投资。45我们指出，电子商务是在 eBay 和亚马逊之后发展起来的，这些由风险投资支持的公司表现非常出色。作者兼天使投资人杰里纽曼指出，仅在 1970 年至 1983 年的五年间，风险投资就增长了 16 倍，从 2.18 亿美元增至 26 亿美元。46这些资金来自有限合伙人——公司、州、主权财富基金，尤其是大型养老基金——这些有限合伙人又投资于风险投资公司。关于这些有限合伙人是否进行了明智的投资，记录好坏参半：正如汤姆·尼古拉斯在《风险投资：美国历史》一书中指出的那样，总体而言，风险投资的回报并不比其他形式的投资好多少。47尽管如此，个人风险投资家往往认为他们才是会做对风险投资家。在过去二十年里，风险投资一直是数据驱动的注意力经济的重要组成部分，包括广告技术公司以及面向消费者的新闻初创公司，如 Vice 和 BuzzFeed。
Has venture capital been with us forever? Although investments have been, many people point to World War II as a time when, among other things, the venture model was born.44 As an example, in 1946, George Doriot, later a Harvard Business School professor, was the quartermaster general in charge of logistics during World War II. After the war Doriot created ARDC, a publicly traded company for investing in long-term research developments, including many developments in the nascent computing industry. In subsequent decades microprocessors and personal computing were to a great extent funded initially through venture capital; given the positive changes of, for example, personal computing on society and the economy, VC has been widely celebrated. Absent from this celebratory timeline is decades of military funding of computation, threaded through earlier chapters, as well as massive support for innovative small businesses through SBIR programs whereby the government effectively served as venture capital.45 E-commerce, we’ve pointed out, grew in the wake of eBay and Amazon, VC-backed companies that did very well. The author and angel investor Jerry Neuman points out a sixteen-fold rise just over the five years 1970 to 1983 in venture capital, from $218 million to $2.6 billion.46 This is funded by limited partners—companies, states, sovereign wealth funds, and especially huge pension funds—which, in turn, invest in the VC firms. The record is mixed as to whether these limited partners are investing wisely: as Tom Nicholas points out in VC: An American History, in aggregate, venture capital doesn’t return much better than other forms of investing.47 Nevertheless individual venture capitalists often believe that they are the venture capitalists who are going to do it right. Venture capital has been a big part of the data driven attention economy in the last two decades, including in advertising technology companies as well as consumer-facing news start-ups such as Vice and BuzzFeed.
风险投资降低风险：投资者提供丰富的资源，使初创公司能够寻找可重复、可扩展的商业模式。需要明确的是，风险仍然是该模式固有的：风险投资家预计其投资组合中的绝大多数公司都会失败，只要少数成功的公司能带来巨额的意外之财，以弥补投资组合中其他公司的投资损失。在许多著名的案例中，例如仙童半导体公司的贡献随着计算技术的兴起，风险是一种技术风险，例如“能否用硅片廉价可靠地制造出集成电路？”风险投资可以避免的另一个风险是市场风险：找到产品与市场的契合意味着创造一种人们愿意以公司可以维持的价格购买的产品。然而，风险投资可以让公司在决定收入模式之前就找到新用户。例如，Facebook 和谷歌就采取了这种做法，它们在确定收入模式并将其强加给用户之前，就积累了庞大的用户群——人们愿意将产品融入他们的生活和习惯。
Venture capital de-risks: investors provide abundant resources that enable start-ups to execute a search for a repeatable, scalable business model. To be clear, risk remains inherent in the model: VCs expect the vast majority of their portfolio companies to fail, so long as the few that succeed return such a large financial windfall as to make up for the lost investment in the rest of the portfolio. In many celebrated cases, such as Fairchild Semiconductor’s contribution to the rise of computing, the risk is a technical risk, such as, “can an integrated circuit be created cheaply and reliably from silicon?” A separate risk which VC can obviate is market risk: finding product-market fit means creating a product people will pay for at a price which allows a company to sustain. However, VC investments allow a company to grow to find new users even before deciding on a revenue model. This was the path taken, for example, by Facebook and Google, which grew a huge user base—people willing to integrate the product into their lives and habits—before settling on a revenue model and imposing it on their users.
最近，风险投资家主张进行规模投资，即所谓的闪电式扩张，即公司通过收购整个市场，从尚未盈利的状态过渡到占据市场主导地位。这种策略不是与市场上的其他公司竞争，而是提供足够的投资来收购市场。例如，Uber 可以以低于市场价格收购现有的整个出租车和豪华轿车市场，不受为驾驶员和乘客提供安全保护的法规的约束。同样，共享办公公司 WeWork 试图向各个城市提供廉价灵活的办公空间，其规模将超越那些依赖于从消费者那里获得可观收入的竞争对手。据Fast Company报道，风险投资家孙正义 (Masayoshi Son) 进行了闪电式扩张，没有进行广泛的尽职调查就投资了创始人兼首席执行官亚当·诺伊曼 (Adam Neumann)：
More recently, venture capitalists have advocated for a scale of investing—so called blitzscaling—which allows companies to attempt to go from pre-revenue to market dominance, effectively by purchasing an entire market. Rather than competing with other companies in the marketplace, this strategy is to provide sufficient investment simply to buy the market. As an example, Uber could afford to undercut entire existing taxi and limousine markets, municipality by municipality, unencumbered by regulations which provided driver and passenger safety protections. Similarly, the coworking company WeWork attempted to flood cities with cheap and flexible office spaces, at a scale which would out-compete competitors reliant on a profitable scale of revenue from consumers. According to Fast Company, the venture capitalist Masayoshi Son blitzscaled, investing in the founding CEO Adam Neumann, without extensive diligence:
坐在后座上的孙正义拿出一台 iPad，写下了对这家公司 44 亿美元投资的条款。他在底部画了两条水平线，在其中一条上签上了自己的名字，然后把 iPad 递给当时 37 岁的诺依曼，让他在另一条上签上自己的名字。诺依曼会手机里留着协议的照片。“当 Masa 第一次选择投资我时，他只见了我 28 分钟。好吗？” 48
There, in the back seat, Son took out an iPad and wrote out the terms for a $4.4 billion investment in the company. He drew two horizontal lines at the bottom, signed his name across one, and then handed the iPad to the then 37-year-old Neumann to scribble his name on the other. Neumann would keep a photo of the agreement on his phone. “When Masa chose to invest in me for the first time, he only met me for 28 minutes. Okay?”48
垄断可以通过购买获得，但也可以通过强化效应和各种循环来发展。数据赋能公司通常会为此提供一种特定的模式。用风险投资家李开复的话来说，“更多的数据会带来更好的产品，进而吸引更多的用户，用户会产生更多的数据，从而进一步改进产品。数据和现金的结合也吸引了顶尖的人工智能人才加入顶尖公司，扩大了行业领先者和落后者之间的差距。” 49这种模式加上投资，帮助谷歌、Facebook 等公司远远优于竞争对手——这要归功于用于训练机器学习算法的大量数据——以至于它们分别在搜索和社交媒体的当前市场上占据主导地位。
Monopolies can be bought, but can also be grown through reinforcing effects and various cycles. Often data-empowered companies provide a particular model for this. In the language of the venture capitalist Kai-Fu Lee, “more data leads to better products, which in turn attract more users, who generate more data that further improves the product. That combination of data and cash also attracts the top AI talent to the top companies, widening the gap between industry leaders and laggards.”49 This model, along with investment, has helped make services such as Google, Facebook, and others so much better than their competitors—owing to a tremendous abundance of data used to train machine learning algorithms—that they clearly dominate the present markets in search and social media, respectively.
注意力经济和风险投资的后果
Consequences of Attention Economy and Venture Capital
上文中我们指出，对注意力经济的担忧至少已有五十年历史，风险投资可以追溯到近七十五年前，而公共关系则伴随我们至少有一个世纪。然而，优化的计算影响力与当代风险投资提供的快速规模相结合是一种特别有效的组合，我们仍在学习将其融入我们的政治和个人现实中。举例来说：Facebook 的首任 CTO 早在 2006 年就创建了一个名为“增长”的子团队——这对于理解 Facebook 的运作方式至关重要。他们是一支由一些最优秀的工程师组成的团队，擅长确保新服务的增长，以越来越多的活跃用户。要做到这一点，就需要确定关键绩效指标 (KPI)，并利用优化这些 KPI 所需的数据和计算基础设施。Facebook 副总裁 Andrew Bosworth 在 2016 年 6 月的一份备忘录中写道：“丑陋的事实是，我们如此坚信要将人们紧密联系在一起，以至于任何能让我们更频繁地联系更多人的事情都是*事实上*的好事情。就我们而言，这也许是指标唯一能说明真实情况的领域。” 50其他工程师，包括那些曾在信息平台公司工作的工程师，也指出了其中的问题。曾在 YouTube 工作的工程师 Guillaume Chaslot 写道：“观看时间是首要任务。其他一切都被视为干扰。” 51不惜一切代价实现增长对 CEO 来说是好事，但对整个行业来说并不总是好事。在《经济的本质》一书中，Jane Jacobs 警告说，公司的增长就像生态系统中的物种一样，可能会以破坏生态系统本身的方式增长。当损害的是市场和社会生态系统本身时，有什么权力可以阻止这种增长？52
Above we’ve argued that concerns about the attention economy are at least fifty years old, that venture capital can be traced back almost seventy-five years, and public relations has been with us for at least a century. The combination of optimized computational influence and the rapid scale provided by contemporary venture capital, however, is a particularly potent mix that we are still learning to integrate into our political and personal realities. To illustrate: the very first CTO of Facebook already created a sub-team in 2006 called “growth”—essential for understanding how Facebook works. They are a team of some of the best engineers, excelling at ensuring the growth of new services to more and more active users. Doing so requires identifying key performance indicators (KPIs) and leveraging the data and computational infrastructure needed to optimize these KPIs. Facebook VP Andrew Bosworth wrote in a June 2016 memo: “The ugly truth is that we believe in connecting people so deeply that anything that allows us to connect more people more often is *de facto* good. It is perhaps the only area where the metrics do tell the true story as far as we are concerned.”50 Other engineers, including those who are formerly at information platform companies, have pointed out problems there. The engineer Guillaume Chaslot, formerly of YouTube, writes, “watch time was a priority. Everything else is considered a distraction.”51 Growth at all cost is good for the CEO, but not always good for the arena. In The Nature of Economies, Jane Jacobs warns of the growth of companies that, like species in an ecosystem, can grow in ways that damage the ecosystem itself. What powers exist to check such growth when the damage is to the marketplace and the societal ecosystem itself?52
* Cory Doctorow 写道：“监控资本家就像舞台上的心理学家，他们声称自己对人类行为的非凡洞察力让他们能猜出你写下并折叠在口袋里的单词，但实际上他们使用托儿、隐藏摄像头、手法和强力记忆来让你惊叹。”“如何摧毁‘监控资本主义’”， OneZero（博客），2020 年 8 月 26 日，https://onezero.medium.com/how-to-destroy-surveillance-capitalism-8135e6744d59；参见 Tim Hwang，《次级注意力危机：广告和互联网核心的定时炸弹》（Farrar, Straus & Giroux，2020 年）。
* Cory Doctorow writes, “Surveillance capitalists are like stage mental-ists who claim that their extraordinary insights into human behavior let them guess the word that you wrote down and folded up in your pocket but who really use shills, hidden cameras, sleight of hand, and brute-force memorization to amaze you.” “How to Destroy ‘Surveillance Capitalism,’ ” OneZero (blog), August 26, 2020, https://onezero.medium.com/how-to-destroy-surveillance-capitalism-8135e6744d59; see Tim Hwang, Subprime Attention Crisis: Advertising and the Time Bomb at the Heart of the Internet (Farrar, Straus & Giroux, 2020).
本书的目的是通过历史的视角来理解数据。本章将我们带入数据的未来。从长远来看，预测未来是一种大胆但具有推测性的方式；在这种情况下，我们将站在更坚实的立场上，预测当前的竞争将如何影响不久的将来。简而言之，我们要问：目前大国之间的竞争是什么，这些竞争的解决将有助于决定数据的未来？除了数据和真相之外，数据和权力一直是我们不变的主题。
The goal of this book has been to understand data through the lens of history. This chapter brings us to the future of data. The brave but speculative way to cover the future, long term, is to predict it; in this case we’re going to stand on firmer ground, by predicting how present contests may shape the near future. In short we ask: What are the present contests among powers whose resolution will help determine the future of data? Alongside data and truth, data and power has been a constant theme of ours.
除了社会的另外两个轨道——政府和企业之外，公民社会也能发挥重要作用。
Civil society can also play an important role [in addition to] the other two rails of society— government and business.
—Karl Manheim 和 Lyric Kaplan 1
—Karl Manheim and Lyric Kaplan1
在思考权力的各种方式中，从米歇尔·福柯的视角到戈登·盖柯的视角，我们反而借鉴了不稳定的三人游戏的隐喻，这种隐喻受到威廉·珍妮薇对三人游戏的描述所启发。推动技术创新的权力，尽管超越了经济权力，更广泛地描述了权力：我们关注企业权力、国家权力和人民权力之间的可变关系。2关于这个比喻，这场不稳定的三人游戏的赢家尚不可预测。要使数据与民主兼容，使数据服务于创造一个公正和繁荣的社会，就需要在这些权力之间找到一种一致性，使公民能够而不是被剥夺权利，增进正义，帮助克服权力差距而不是巩固权力差距。
Of all the various ways to think about power, from the lens of Michel Foucault to that of Gordon Gekko, we draw, instead, on the metaphor of an unstable three-player game, inspired by William Janeway’s account of the three forces driving technology innovation, though expanded beyond economic forces to describe power more generally: we focus on the mutable relations among corporate power, state power, and people power.2 Apropos of this metaphor, the winners of this unstable three-player game are not yet predictable. Making data compatible with democracy, and making data serve the creation of a just and flourishing society, will require finding a conformation among these powers that enables rather than disables citizens, that enhances justice, that aids in overcoming power disparities rather than entrenching them.
权力第一：企业权力
Power the First: Corporate Power
我们关注的是数据和数据赋能算法的使用所引发的技术和社会技术问题；本书的最后一部分集中讨论了企业权力，特别是目前占主导地位的大型科技公司。我们很自然地会问，为什么制造这些问题的公司和技术人员不继续解决这些问题呢？我们可能会更微妙地问，考虑到问题的范围，即使他们想解决所有问题，他们能做到吗？目前尚不清楚公司是否有动力改变他们与数据的关系，至少盈利的公司没有。即使在 Web 2.0 时代之前，许多公司在隐私方面的记录也好坏参半。2010 年，Facebook 首席执行官马克·扎克伯格吹嘘道：
We’ve focused on technological and socio-technological concerns raised by the use of data and data-empowered algorithms; the final part of the book has centered on corporate power, particularly in the currently dominant big tech companies. It’s natural to ask, Why don’t the companies and the technologists who created these problems go ahead and fix them as well? Slightly more subtly we might ask, Given the scope of the problems, even if they wanted to fix everything, could they? It’s not clear that companies are motivated to change their relationship with data, at least not the profitable ones. Even before the age of Web 2.0, many companies had a mixed record on privacy. In 2010 Facebook CEO Mark Zuckerberg bragged:
许多公司会受困于惯例和他们所建立的遗产，无法进行隐私变更——为 3.5 亿用户进行隐私变更并不是许多公司会做的事情。但我们认为这是一件非常重要的事情，要始终保持初学者的心态如果我们现在创办一家公司，并且我们决定这些将成为现在的社会规范，那么我们就会这样做。3
A lot of companies would be trapped by the conventions and their legacies of what they’ve built, doing a privacy change—doing a privacy change for 350 million users is not the kind of thing that a lot of companies would do. But we viewed that as a really important thing, to always keep a beginner’s mind and what would we do if we were starting the company now and we decided that these would be the social norms now and we just went for it.3
他们确实这么做了。然而，即使开关控制着近 30 亿（截至 2022 年 1 月）用户，规范也不一定会通过开关的翻转做出反应。相比之下，可能是为了应对这种情况，一些公司不仅将消费者保护作为价值主张，而且作为竞争优势。苹果首席执行官蒂姆·库克 (Tim Cook) 在 2015 年宣布“隐私是一项基本人权”——这是一种消费者保护立场，也是强有力的营销文案，现在当您首次启动许多苹果产品时就会看到它。4由于苹果仍然主要是一家硬件公司，而不是广告公司，因此它不会面临信息平台公司所面临的消费者隐私和收入之间的直接冲突，后者的主要收入来自定向广告。
Indeed they did. Norms, though, don’t necessarily respond with the flip of a switch, even when that switch governs nearly three billion (as of January 2022) users. By contrast, and possibly in response, some companies have taken on consumer protection not only as a value proposition but as a competitive advantage. Tim Cook, CEO of Apple, declared “privacy is a fundamental human right” in 2015—a consumer protection position, and powerful marketing copy that now greets you when you first boot up many Apple products.4 As Apple remains predominantly a hardware company rather than an advertising company, it does not face the direct tension between consumer privacy and revenue experienced by information platform companies whose primary revenue is via targeted advertising.
道德即服务
ETHICS AS A SERVICE
最近，几家大公司已经采取行动，表明它们可以为我们所概述的道德问题提供技术解决方案。其中一些是内部努力，即通过社会技术尝试使公司的政策与道德原则保持一致；其他则是外部努力，包括计算工具集和咨询服务，以帮助研究人员和其他公司将道德付诸实践。一个引人注目的内部努力的例子是谷歌的道德人工智能团队，我们之前讨论过该团队的解散。
Recently several major companies have made moves to show that they can provide technical solutions to the ethical problems we’ve outlined. Some of these are internal efforts, i.e., socio-technical attempts to align the company’s policies with ethical principles; others are external efforts, both computational toolsets as well as consulting services, to help researchers and other companies bring ethics into practice. An attention-getting example of an internal effort was Google’s Ethical AI team, whose implosion we discussed earlier.
内部与外部企业道德规范
INTERNAL VS. EXTERNAL CORPORATE ETHICS-MAKING
外部企业道德服务应与内部道德制定形成对比。如上所述，内部努力包括创建道德 AI 团队谷歌负责研究人工智能实践的伦理影响，包括谷歌其他研究小组的实践。从组织权力动态的角度来看，直接批评公司内部的工作可能极其困难，当这项工作与收入来源有关时，就更是如此。一位前谷歌软件工程师总结了这样一个团队及其主管玛格丽特·米切尔博士的职能：
External corporate ethics as a service should be contrasted with internal ethics-making. Internal efforts include, as discussed above, the creation of the Ethical AI team within Google, charged with researching the ethical impact of AI practices including those of other Google research groups. It can be extremely difficult, in terms of organizational power dynamics, directly to critique work done within a company, even more so when that work is related to sources of revenue. A former Google software engineer summarized the function of such a team and its director, Dr. Margaret Mitchell:
我和他们交流时，就好像他们是一群来咨询的专家。事实上，她（米切尔）正在构建一个模型，每个人工智能团队都应该如何工作，并将道德作为技术发展的首要关注点。5
I had interacted with them as though they were a group of experts who would come in to consult. In fact what she [Mitchell] was building was a model of how every single AI team should work, with a mind towards ethics as a primary concern of technological development.5
他后来将道德团队描述为“遮羞布”，其影响次于“季度收益”；6达纳·博伊德 (Danah Boyd) 在 2016 年引用诗人兼教授奥德丽·洛德 (Audre Lorde) 的话，表达得更加诗意：
He would later describe the ethics team as “a fig leaf” whose impact was made secondary to “quarterly earnings;”6 danah boyd put it more poetically in 2016, quoting poet and professor Audre Lorde:
虽然我们认为自己了解战争和心理实验的伦理，但我认为我们对于如何真正管理组织中的伦理却一无所知。……奥德丽·洛德 (Audre Lorde) 说过，“主人的工具永远不会拆毁主人的房子。”从某种意义上说，我同意。但我也看不出有什么方法可以向一个复杂的系统扔石头来实现伦理。7
While we think we understand the ethics of warfare and psychology experiments, I don’t think we have the foggiest clue how to truly manage ethics in organizations. . . . Audre Lorde said, “the master’s tools will never dismantle the master’s house.” And, in some senses, I agree. But I also can’t see a way of throwing rocks at a complex system that would enable ethics.7
简而言之，通过企业自我批评整合道德规范的挑战尚未解决。在大学里，机构审查委员会通过控制财政大权来控制权力，并可以决定研究的道德性。企业没有明显的类似物：具有真正权力的内部结构，无论是金融的还是其他的。
In short, the challenge of integrating ethics via corporate self-critique has not yet been solved. In universities, institutional review boards control power by controlling the purse strings and can dictate ethical research. Corporations have no obvious analogue: internal structures with real teeth, financial and otherwise.
另一方面，外部企业道德没有那么复杂的政治因素，并允许科技公司向前发展，特别是隐含的论点，即道德从根本上来说是一个技术问题，最好用技术解决方案来解决。最近的例子包括：
External corporate ethics, on the other hand, have less complicated politics and allow tech companies to advance, in particular, the implicit thesis that ethics is fundamentally a technical problem best met with technical solutions. Recent examples include:
• IBM 创建了“AI Fairness 360”，这是一个具有“9 种算法和多种指标”的开源工具包；
• IBM created “AI Fairness 360,” an open-source tool kit with “9 algorithms and many metrics”;
• 谷歌发布了“What-If Tool”以及涉及公平方面的“Facets”；
• Google released the “What-If Tool” along with “Facets,” which concerns facets of fairness;
• 微软有自己的学习工具集“Fairlearn”（一个 Python 模块）；
• Microsoft has its own tool set for learning, “Fairlearn” (a Python module); and
• Facebook 有自己的工具集，称为“Fairness Flow”。
• Facebook has its own tool set called “Fairness Flow.”
咨询公司埃森哲也开发了自己的工具来消除算法中的偏见。咨询公司有很多机会帮助其他公司制定和体现科技道德。例如，数据科学家兼作家 Cathy O'Neil 就拥有一家咨询公司来做这件事；Rumman Chowdhury 是埃森哲最直言不讳的科技伦理学家之一，他于 2020 年短暂离开公司，创建了一家基于工具的咨询公司 Parity，之后加入 Twitter 领导其（现已解散的）团队 META（机器学习、道德、透明度和问责制）。8当然，大型科技公司也能享受到咨询的成果，例如谷歌在 2020 年 8 月宣布正在探索科技道德咨询服务，几个月后，它解雇了道德 AI 团队的联合创始人。
Accenture, a consulting company, has also developed its own tools to eliminate bias in algorithms; there are plenty of opportunities for consulting companies to help other companies frame and manifest tech ethics. The data scientist and author Cathy O’Neil, for example, has a consulting company for doing so; Rumman Chowdhury, one of the most outspoken of Accenture’s tech ethicists, left the company to create a tool-based consulting company, Parity, briefly in 2020 before joining Twitter to lead their (now dismissed) team META (for Machine Learning, Ethics, Transparency and Accountability).8 Of course large tech companies also enjoy the fruits of consulting, as when Google announced in August 2020 that it was exploring tech ethical consulting as a service, a few months before firing the cofounders of the Ethical AI team.
所有这些都导致科技公司在道德方面，尤其是在公平的技术方法方面，对自身定位不一。但这隐含着对问题的重新定义当公司推进技术解决方案时，例如上述公平工具。我们强调，要理解数据赋能的产品和服务如何违反我们的规范和价值观，需要广泛的社会技术视角。在这种社会技术复杂性中，存在着应用伦理学的主题；在应用伦理学中，存在着正义；在正义中，存在着公平；在公平中，存在着公平的量化；在公平与个人的其他利益和组织的目标之间的平衡。对于技术人员来说，提升技术水平是很自然的。举个例子，计算机科学家 Michael Kearns 和 Aaron Roth 在《道德算法》（2019 年）中写道：
All of these contribute to a checkered positioning of technology companies that they are “on it” with respect to ethics, and in particular with respect to technical approaches to fairness. But there’s an implicit reframing of the problem when companies advance technological solutions, such as the fairness tooling above. We’ve emphasized that understanding how data-empowered products and services violate our norms and values requires a broad, socio-technical view. Within this socio-technical complexity lies the topic of applied ethics; and within this, justice; and within justice, fairness; and within fairness, the quantification of fairness; and within this, the balance between fairness and other interests of individuals and goals of organizations. Elevating the technical is natural for technologists. As one example, the computer scientists Michael Kearns and Aaron Roth write in The Ethical Algorithm (2019):
我们……相信，遏制算法不当行为本身就需要更多更好的算法——这些算法可以协助监管机构、监督组织和其他人类组织监控和衡量机器学习的不良和非预期影响。9
We . . . believe that curtailing algorithmic misbehavior will itself require more and better algorithms— algorithms that can assist regulators, watchdog groups, and other human organizations to monitor and measure the undesirable and unintended effects of machine learning.9
这种观点将问题分解为技术要素，例如“算法”将“协助”社会要素，例如“人类组织”，例如“监督团体”。事实上，机器学习已被证明是一种强大的方法，可与复杂环境相结合进行统计和计算优化。机器学习专家迈克尔·乔丹和汤姆·米切尔写道：“从概念上讲，机器学习算法可以看作是在训练经验的指导下，在大量候选程序中搜索，以找到优化性能指标的程序。” 10通常，该指标表示准确性和复杂性之间的权衡，但如果我们最大限度地提高准确性和公平性，同样的方法也会起作用。“这个优化问题的版本”，计算机科学家辛西娅·鲁丁写道，“是人工智能的一些基本问题”。将道德重新定义为准确性（或利润或其他量化目标）与复杂性（或邪恶或某些可量化代理）之间的权衡，仍然需要有人指定权衡。在避免模型复杂性的背景下，将这样的参数设置为 1% 意味着，正如鲁丁所解释的那样，“我们会牺牲 1% 的训练准确率来将模型大小缩小一倍。” 11公平的技术领域吸引了来自这种统计作为计算优化框架的技术人员，该框架主导了我们在 21 世纪对“机器学习”的定义。虽然它可以改善算法系统的某些方面，但这种技术解决方案往往回避了这些系统的权力和社会嵌入的更大问题，正如海伦·尼森鲍姆和萨菲亚·诺布尔等批评家多年来所记录的那样。
This view parses the problem into the technical elements, such as “algorithms” which will “assist” the societal elements, such as “human organizations,” such as “watchdog groups.” Indeed, machine learning has proven a powerful approach for statistical and computational optimization coupled to complex environments. “Conceptually,” the machine learning experts Michael Jordan and Tom Mitchell write, “machine-l earning algorithms can be viewed as searching through a large space of candidate programs, guided by training experience, to find a program that optimizes the performance metric.”10 Often this metric expresses a tradeoff between accuracy and complexity, though the same methods work if we instead maximize a combination of accuracy and, for example, fairness. “Versions of this optimization problem,” writes computer scientist Cynthia Rudin, “are some of the fundamental problems of artificial intelligence.” Reframing ethics as a trade-off between accuracy (or profit, or some other quantified goal) and complexity (or evil, or some quantifiable proxy) still requires someone to specify the tradeoff. In the context of avoiding model complexity, setting such a parameter to 1 percent means, as Rudin explains, “we would sacrifice 1% training accuracy in order to reduce the size of the model by one.”11 The technical field of fairness appeals to technologists coming from this statistics-as-computational-optimization framework, which dominates what we mean by “machine learning” in the twenty-first century. And while it can ameliorate some facets of algorithmic systems, such technological solutions often sidestep bigger questions of the power and social embedding of those systems, as critics from Helen Nissenbaum to Safiya Noble have documented for years.
威胁和误导
THREATS AND MISDIRECTION
将对人工智能危险的讨论缩小到可以通过改变优化算法来解决的危险，这是对问题本质讨论的短视之一。其次是让未来主义的科幻梦想和噩梦占据主导地位，诱使我们忽视现有系统的当前问题和危害，将注意力引向超智能人工智能和“通用人工智能”（GAI）的推测未来状态，而后者的惊人无所不能需要当代的辩护。这种特殊厄运的预言者包括硅谷一些最根深蒂固的思想领袖，如特斯拉的埃隆马斯克或谷歌的雷库兹韦尔。然而，正如 Annette Zimmermann、Elena Di Rosa 和 Hochan Kim 指出的那样：
Narrowing the discussion of the perils of AI to those that can be addressed by changing the optimization algorithm is one myopia of discussions on the nature of the problem. A second is allowing futurist sci-fi dreams and nightmares to dominate, enticing us to ignore present concerns and harms of existing systems by drawing attention to the conjectured future state of hyper-intelligent AI and “general AI” (GAI) whose awesome omnipotence demands a contemporary defense. Prophets of this particular doom include some of Silicon Valley’s most embedded thought leaders, such as Tesla’s Elon Musk or Google’s Ray Kurzweil. However, as Annette Zimmermann, Elena Di Rosa, and Hochan Kim point out:
别在意遥远的末日幽灵；人工智能已经存在，在许多事情的幕后工作。我们的社会制度……我们必须抵制关于人工智能的充满末日色彩的言论，这种言论会鼓励习得性无助的心态。12
Never mind the far-off specter of doomsday; AI is already here, working behind the scenes of many of our social systems. . . . We must resist the apocalypse-saturated discourse on AI that encourages a mentality of learned helplessness.12
作者认为，终结者机器人的闪亮幽灵不应该转移人们对当前人类决策（包括企业产品开发商的决策）在带来当前对权利、正义和民主的危害和挑战方面所起的作用的注意力。
The shiny specter of terminator robots, these authors argue, shouldn’t distract from the role that present human decisions, including those made by corporate product developers, play in bringing about current harms and challenges to rights, justice, and democracy.
开发算法系统需要做出许多深思熟虑的选择。。。算法本身并不定义这些概念；人类——开发人员和数据科学家——选择诉诸哪些概念，至少作为最初的起点。13
Developing algorithmic systems entails making many deliberate choices. . . . The algorithm does not define these concepts itself; human beings—developers and data scientists—choose which concepts to appeal to, at least as an initial starting point.13
在这里，齐默曼、迪罗莎和金呼应了吉娜·内夫及其合著者的观点，即数据科学的实践，特别是在开发和部署机器学习产品时，涉及无数主观的设计选择，每一个都是反思和批评的机会，无论是内部还是外部。14使用“数据”或“算法”等术语并不能将这类工作从其主观性和政治性中解放出来。更广泛地说，我们不能让首席执行官或企业传播部门为我们定义人工智能的哪些威胁最应该引起我们的注意。
Here, Zimmermann, Di Rosa, and Kim echo a point made by Gina Neff and coauthors that the practice of data science, in particular when developing and deploying machine learning products, involves innumerable subjective design choices, each of which is an opportunity for reflection and critique, internal and external.14 The use of the terms “data” or “algorithm” does not liberate such work from its subjectivity and its politics. More broadly, we cannot let CEOs or corporate communications departments define for us what threats about AI should most draw our attention.
企业去平台化和企业联盟
CORPORATE DEPLATFORMING AND CORPORATE COALITIONS
消费者对企业使用数据赋能算法的担忧日益增加，与此同时，企业的举动也随之而来。尽管这些举动可能以保护消费者的名义出现，但更准确地说，应该是以保护消费者的名义进行的企业间竞争。其中一些竞争是当今最强大的科技公司之间的较量。还有一些竞争不仅涉及巨头之间的一对一较量，还涉及更为复杂的联盟，这些公司组成不断变化的联盟，扮演着各种不同的角色，如数据提供商、搜索引擎、销售线索生成器，或者在移动应用商店的情况下，充当能够阻挡用户并进而阻挡其他公司收入的关键瓶颈。
The rise of consumer concerns regarding corporate use of data-empowered algorithms has coincided with corporate moves which, though they may be presented as consumer protection, may be more accurately characterized as intercorporate competitions in the name of consumer protection. Some of these contests pit the most powerful tech companies today against each other. Others involve not just one-on-one matchups between titans but more complicated coalitions, in which companies form shifting alliances playing such diverse roles as data providers, search engines, lead generators for sales, or, in the case of mobile app stores, key choke points capable of blocking the users and thereby the revenue of other companies.
随着私营公司逐渐形成技术生态系统的必要基础设施，它们相互依赖，从而相互胁迫。例如，苹果公司就拥有惊人的影响力，可以影响那些希望其应用在 iPhone 上可用的其他公司。在《Super Pumped》一书中，迈克·艾萨克描述了优步当时的首席执行官拜访苹果首席执行官蒂姆·库克的一次经历，库克毫不含糊地解释了如果优步想留在应用商店，该应用将允许哪些数据政策。15同样，Facebook 和其他数据驱动的公司希望留在苹果的应用商店中，因此应用商店的功能就像基础设施一样。实际上，这些公司可以相互剥夺对方的平台，例如当谷歌通过改变搜索算法来惩罚一家公司，从而损害该公司的搜索排名，导致巨额财务损失时。这些公司有机会相互影响，既可以通过定位自己，例如当苹果公司将隐私视为一项基本人权来抨击谷歌和 Facebook 时，也可以通过实际剥夺对方的平台。一家公司加强隐私的改变可能会导致其他公司损失数十亿美元。16在我们的权力分析中，这些代表了企业权力中的一组重要箭头。
As private companies have come to form the necessary infrastructure for technology ecosystems, they rely on each other—and thereby coerce each other. Apple has, for example, shocking power to influence other companies that wish to have their app available on the iPhone. In Super Pumped, Mike Isaac describes a visit by the then-CEO of Uber to Tim Cook, the CEO of Apple, who explains in no uncertain terms what data policies are going to be allowed in the app if Uber wants to remain in the App Store.15 Similarly, Facebook and other data-enabled companies wish to remain in Apple’s app store, which therefore functions like infrastructure. Effectively, these companies can deplatform each other, as when Google punishes a company by changing the search algorithm in ways that hurt that company’s search ranking, leading to a huge financial loss. These corporations have the opportunity to affect each other, both by positioning themselves, as when Apple celebrates privacy as a fundamental human right to poke at Google and Facebook, and by actually deplatforming each other. Privacy-enhancing changes at one company can lead to billions of dollars in loss at others.16 In our analysis of power these represent an important set of arrows among corporate powers.
有些竞争涉及复杂且不断变化的企业权力联盟。十年前的一个例子是，不同行业的不同部分与华盛顿一些传统上最强大的游说团体相匹配：娱乐业（尤其是美国电影协会和华特迪士尼公司）发现自己与大量所谓的内容平台对立。这场斗争围绕着 2012 年的一项名为《禁止网络盗版法案》（SOPA）的法案展开，该法案设想大幅扩大对各种内容（包括流媒体）的知识产权执法。2012 年 1 月 18 日，一个科技公司联盟与电子前沿基金会等公民自由组织合作，上演了一场虚拟停电。例如，谷歌在搜索登录页面上审查了自己的公司名称。在公众看来，这是一个由利益各异但又足够共同的人组成的临时联盟，他们反对该法案。17
Some contests involve complex, shifting alliances of corporate powers. An example from a decade ago matched different parts of industry with some of the most traditionally powerful lobbies in Washington: the entertainment industry (in particular, the Motion Picture Association along with The Walt Disney Company) found themselves arrayed against a large number of the so-called content platforms. This became a fight around a bill in 2012 called the Stop Online Piracy Act (SOPA), which envisioned a dramatic expansion of intellectual property enforcement around a variety of content, including streaming media. On January 18, 2012, a coalition of tech firms collaborating, unusually, with civil liberties organizations such as the Electronic Frontier Foundation, staged a kind of virtual blackout. Google, for example, censored its own corporate name on the search landing page. What appeared to the public was an ad hoc coalition of people with interests which were diverse yet common enough to push back against the bill.17
可以理解，人们对于各个相关组织的动机持怀疑态度，但正是这种由利益各异的人组成的联盟能够带来各种变化。苹果和其他公司已经推出了各种创新来增强我们的隐私；微软研究院的差异隐私就是其中之一。举个例子：苹果公司与美国司法部就联邦政府要求他们做什么才能解锁 iPhone 进行了旷日持久的斗争，而他们已经为 iPhone 提供了前所未有的保护。不同的联盟可以将企业权力引导到特定的方向，将企业能力与更民主的权力结合起来。通过复杂的团结，强有力的行动不需要完全达成共识。
It’s understandable to be cynical about the motives of the various organizations involved, yet it’s precisely these kinds of coalitions of people with diverse interests that can affect kinds of change. Apple and other companies have introduced a variety of innovations which enhance our privacy; differential privacy, arising from Microsoft Research, is another. An example: Apple has been in a protracted battle with the US Department of Justice over what they are required to do by the federal government in order to unlock iPhones, which they have given an unprecedented level of protection. Diverse coalitions can channel corporate power in particular directions that combine corporate capacities with more democratized power. Total consensus is not a requirement for potent action made possible by complex solidarities.
自律组织
SELF-REGULATORY ORGANIZATIONS
企业权力对国家权力崛起的一种反应是建立和推广自律组织 (SRO)。自律组织对数据来说并不陌生：该术语起源于 20 世纪 30 年代的证券法和改革。目前，尽管缺乏实际的监管职能，但这一术语描述了各种各样的此类伙伴关系；也就是说，这一术语指的是不进行严格监管，而是进行研究、召开会议并撰写报告以解释甚至批评相关公司工作的组织。尽管在治理上独立，但它们在财务上往往依赖于它们研究的公司。企业人工智能领域的一个突出例子是 2016 年成立的非营利组织人工智能伙伴关系。此类团体模糊了公民社会（包括教育和研究机构）、企业权力和资金以及通常与实际监管相关的监管制衡类型（即国家权力）之间的界限。这种批评和资金之间的矛盾关系导致批评人士认为此类组织被其赞助商“俘获”，从而削弱了三方博弈中的任何权力。
One response by corporate power to the rise in state power is the formation and promotion of self-regulatory organizations (SROs). SROs are not new to data: the term originates in securities law and the reforms of the 1930s. At present a broad variety of such partnerships are described by the term, despite their lack of actual regulatory function; that is, the term refers to organizations which do not strictly regulate, but rather research, convene meetings, and write reports that explain and sometimes critique the work of related companies. Although independent in governance, they are often financially dependent on the companies whose work they research. A prominent example from the world of corporate AI is the Partnership on AI, a nonprofit founded in 2016. Such groups blur the lines between civil society (including educational and research institutions), corporate power and funding, and the type of regulatory checks and balances typically associated with actual regulation; namely, with state power. This conflicted relationship between critique and funding has led critics to dismiss such organizations as “captured,” by their sponsors, defanging any power within the three-party game.
第二种权力：国家权力
Power the Second: State Power
随着企业成为关键基础设施，人们常说，它们的权力可以与民族国家相媲美。然而，当我们思考如何最好地应对企业的过度扩张时，我们通常首先想到的是国家权力，这是对企业权力的最佳制约——也是最佳监管。18
As corporations become critical infrastructure, their power is often said to rival that of nation-states. When we think of how best to respond to corporate overreach, nonetheless we often first think of state power as the best check on—the best regulation of—corporate power.18
企业经常将国家对企业权力的限制描绘成阻碍创新的强大障碍。这种纯粹“压制性”的监管观点是不够的，因为国家权力在资金、法规制定和法律方面更积极、更有建设性地塑造企业权力。2020 年，耶鲁大学法学教授 Amy Kapczynski 解释说：
Corporations often portray state limits on corporate power as an imposing barricade against innovation. Such a purely “repressive” vision of regulation is inadequate, for state power shapes corporate power more positively and constructively, in funding, in regulation-making, and in the law. In 2020 Yale law professor Amy Kapczynski explained:
认为谷歌和 Facebook 的运营是在不受法律约束的地区进行的，甚至认为这些公司所期望的——是错误的。它掩盖了这些公司依赖法律来获得权力的程度，以及许多可以修改以增强公共权力的法律决定。19
the view that the operations of Google and Facebook occur in a law-free zone—or even that those companies would so desire—is wrong. It conceals the degree to which these companies rely upon law for their power and the many legal decisions that could be altered to enhance public power.19
因此，国家权力不仅仅是对企业权力的限制。它为企业结构的发展创造了条件。从美国缺乏隐私方面的一般法律，到资本收益或房地产税收处理的特殊性，国家权力使得某些类型的商业模式、某些类型的大规模数据使用以及对民主秩序的某些挑战成为可能。毫无疑问，我们不能想象联邦监管是解决算法弊病的灵丹妙药，即使它包括可能广泛禁止某项技术。在 20 世纪 70 年代消费者隐私保护法规兴起之后，从历史上看，在美国，当联邦监管实施时，它通常采用“部门”方式，即逐个部门而不是逐个能力进行监管——换句话说，限制行业内某项技术的使用，而不是完全禁止面部识别。因此，美国不太可能采用与欧洲《通用数据保护条例》相当的广泛联邦监管。曼海姆和卡普兰抓住了美国方法中的平衡点：
State power, then, is not merely a limit on corporate power. It creates the conditions under which corporate structures develop. From the lack of general laws around privacy in the United States to the particularities of the tax treatment of capital gains or real estate, state power makes possible certain kinds of business models, certain kinds of mass data use, and certain challenges to democratic order. Unquestionably, we cannot imagine federal regulation as a panacea to algorithmic ills, even if it were to include possible broad banning of a technology. Following the rise of consumer privacy protection regulation in the 1970s, historically, in the United States, when federal regulation takes place, it typically does so in a “sectoral” approach, that is regulated business sector by sector, rather than capability by capability—in other words, limiting the use of a technology within industry rather than, say, banning facial recognition altogether. The US is therefore unlikely to adopt a broad federal regulation comparable to Europe’s General Data Protection Regulation. Manheim and Kaplan capture the balances in the US approach:
许多美国企业最初倾向于采用行业方法，以便根据其细微需求量身定制法规。虽然这种模式有一定的合理性，但它也助长了监管俘获、行业游说和隐私滥用，而这些往往都无法通过监管漏洞。20
Many U.S. businesses initially preferred the sectoral approach as to tailor regulations to their nuanced needs. While there is some validity to that model, it also facilitates regulatory capture, industry lobbying, and privacy abuses often falling through regulatory cracks.20
Kearns 和 Roth 同样强调了这种方法存在监管漏洞的风险。21随着新的商业模式形成后，我们必须问：从部门角度看，Facebook 是出版商还是广告公司？不同的司法管辖区对这些决定可能有相互矛盾的解释，例如 Facebook 被迫取消收购 Giphy，理由是英国竞争与市场管理局监管下的反竞争问题。22在美国占主导地位的部门方法下，这些看似学术性的问题的答案具有巨大的监管影响。而竞争对手很容易质疑一个行业以牺牲另一个行业为代价来获取监管的努力，正如我们在 SOPA 版权改革的例子中所看到的那样。正如曼海姆和卡普兰所言，“州和联邦法律‘相互重叠、衔接和矛盾’的拼凑体系”使得通过广泛保护权利的法律变得困难：
Kearns and Roth likewise emphasize the risk of regulatory cracks in this approach.21 As new business models are formed, we must ask: In a sectoral approach, is Facebook a publisher or an advertising company? Different jurisdictions might have conflicting accounts of these decisions, as when Facebook was forced to un-acquire Giphy, based on anticompetitive concerns as regulated by the United Kingdom’s Competition and Markets Authority.22 Under the sectoral approach dominant in the US, the answers to such seemingly academic questions have tremendous regulatory impact. And rival corporations are prone to question the efforts of one industry to capture regulations at the expense of another, as we saw in the example of SOPA copyright reform. The “patchwork system of state and federal laws that ‘overlap, dovetail and contradict one another’” has made it difficult to pass laws that protect rights broadly, as Manheim and Kaplan argue:
“监控资本主义”之所以繁荣，是因为隐私权严重缺乏保护，我们的法律未能跟上技术的步伐。我们上一部重要的联邦隐私法 (ECPA) 是在 1986 年颁布的，比 Facebook、谷歌和 YouTube 都早，甚至比万维网都早。在此期间，数据和人工智能公司发展壮大，如今在经济、公共政策和我们的生活方面拥有不成比例的权力。23
“surveillance capitalism” prospers because privacy rights are grossly under-protected and our laws have failed to keep pace with technology. Our last major federal privacy law (ECPA) was enacted in 1986, before Facebook, before Google and You-Tube, indeed before the World Wide Web. Data and AI companies have grown and flourished in the interim, now commanding disproportionate power over the economy, public policy, and our lives.23
监管型国家的瓦解和潜在重塑
THE DEFANGING OF–AND POTENTIAL REFANGING OF–THE REGULATORY STATE
反垄断监管是联邦监管的一种形式，它涉及多个部门。美国的反垄断监管在 19 世纪末和 20 世纪初蓬勃发展，以应对大型托拉斯，其中最臭名昭著的是 JD 洛克菲勒的标准石油公司，该公司拥有惊人的权力和主导地位市场。24这种不受制约的权力的滥用远远超出了提高消费者价格的范围。然而，到二十世纪末，反垄断监管在很大程度上被重新解释为仅限于企业权力可能与消费者面临的价格上涨有关的情况。25显然，如果你是产品，免费使用服务，只用你的时间和数据付费，那么这种框架就不适用了。
One form of federal regulation that is multi-sectoral is antitrust regulation. Antitrust regulation in the United States blossomed at the end of the nineteenth and beginning of the twentieth century in response to the great trusts, most infamously J. D. Rockefeller’s Standard Oil Company, which enjoyed an amazing accumulation of power and dominance over markets.24 The abuses of that unchecked power went far beyond raising prices for consumers. By the end of the twentieth century, however, antitrust regulation had largely been reinterpreted as limited to cases in which corporate power could be linked to increased prices faced by consumers.25 Clearly, this framing is inapplicable if you are the product, using the service for free, paying only with your time and your data.
“国家权力”不应等同于美国联邦政府的监管。目前，各种国际法规也对个人数据和数据赋权算法的影响提出质疑，限制了塑造我们数字未来的全球活跃科技公司的行动。最近最明显的例子是欧盟的《通用数据保护条例》（GDPR），该条例于 2018 年 5 月 25 日生效。GDPR 从根本上挑战了监控资本主义的商业模式。GDPR 的一项高级原则，即第 22 条规定，欧洲人“有权不受仅基于自动化处理的决定的约束”。26与道德或宪法原则一样，政策制定者、游说者和法院之间的审议工作开始，将这些原则提炼为标准和规则。除其他标准外，GDPR 还列出了许多“数据主体”（即人）的权利，包括“被遗忘权”。27将这一原则转化为政策对企业提出了挑战，要求他们标准化和改进其数据治理；拥有大量与用户相关的不连贯的记录，且用户身份不同，这大大增加了公司识别和删除提出此类请求的个人记录所需的时间（以及成本）。
“State power” should not be equated with US federal government regulation. The impact of personal data and data-empowered algorithms are at present contested by a variety of international regulations as well, limiting the actions of the globally active tech companies shaping our digital futures. The most visible recent example is the EU’s General Data Protection Regulation (GDPR), which became effective on May 25, 2018. The GDPR fundamentally challenges surveillance capitalism as a business model. A high-level principle with GDPR, Article 22, states that Europeans “have the right not to be subject to a decision based solely on automated processing.”26 As with ethical or constitutional principles, deliberative work among policymakers, lobbyists, and the courts then begins, distilling these principles into standards and rules. Among other standards, the GDPR lays out a number of “rights of the data subjects’” (i.e. people) including the “right to be forgotten.”27 Turning this principle into policy has challenged corporations to standardize and improve their data governance; having a myriad of disconnected records pertaining to users, with differing user identification, greatly increases the time (and thus cost) required for a company required to identify and delete records for individuals who make such requests.
美国的监管也同样在州和地方层面进行，通常具有类似的效果，即迫使公司在全球范围内遵守标准，即使这些标准是可执行的仅在其运营的部分地区才允许。从运营上讲，不这样做将需要在每个地区运行单独的系统和流程，这种后勤复杂性几乎不值得通过区域化获得任何额外利润。美国州一级的一个例子是《加州消费者隐私法案》（CCPA），该法案于 2018 年 6 月 28 日生效，有时被称为“加州的 GDPR”。CCPA 的措辞比 GDPR 更精确，在条款或规则上而不是原则上，CCPA 规定了罚款，例如“加州居民和事件罚款 100 至 750 美元，或实际损失，以较高者为准，以及法院认为适当的任何其他救济，但加州总检察长办公室可以选择起诉该公司，而不是允许对其提起民事诉讼”（加州民法典 §1798.150）。考虑到加州居民的数量和对“事件”的解释，这样的罚款可能会给一家以个人数据为商业模式动力的公司带来巨大的财务负担。正如加利福尼亚州曾经在汽车环境法规方面处于领先地位一样，该州正试图在隐私法规方面处于领先地位。
Regulation in the United States takes place as well at the state and local level, often with similar effects of compelling companies to comply globally even with standards enforceable only in some of the regions in which they operate. Operationally, doing otherwise would require separate systems and processes active in each region, a logistical complexity rarely worth any additional profit gained by regionalization. An example at the US state level is the California Consumer Privacy Act (CCPA), which became effective June 28, 2018, sometimes called “California’s GDPR.” Written with more precision than the GDPR, in terms or rules rather than principles, the CCPA specifies fines, for example, of “$100 to $750 per California resident and incident, or actual damages, whichever is greater, and any other relief a court deems proper, subject to an option of the California Attorney General’s Office to prosecute the company instead of allowing civil suits to be brought against it” (Cal. Civ. Code §1798.150). Given the number of California residents and interpretations of “incident,” such fines could be a tremendous financial burden to a company for whom personal data fuels the business model. Just as California once led the way with environmental regulations of automobiles, the state is attempting to lead the way in privacy regulations.
在地方层面，市政当局在监管某些监控技术方面一直处于领先地位。2019 年 7 月，奥克兰通过了一项法令，禁止“获取、获得、保留和访问”面部识别技术。旧金山于 2019 年 5 月通过了类似的法律；马萨诸塞州萨默维尔于 2019 年 6 月通过了类似的法律；明尼阿波利斯于 2021 年 2 月通过了类似的法律。
At a more local level, municipalities have been in the forefront in regulating some surveillance technologies. In July 2019, Oakland passed an ordinance preventing “acquiring, obtaining, retaining and accessing” facial recognition technology. Similar laws have been passed in San Francisco in May 2019; Somerville, Massachusetts, in June 2019; and Minneapolis in February 2021.
这些趋势是国家权力重新调整的一部分，至少在美国是这样，以制衡企业在数据方面的权力。反垄断的广泛、多部门职权范围最近受到了质疑，哥伦比亚大学的 Tim Wu 和 Lina Khan 等倡导者主张一种较旧的“新布兰代斯主义”观点，这种观点涵盖了对经济集中的危险的更广泛理解：垄断或近乎垄断。28（路易斯·布兰代斯的观点经常与罗伯特·博克的观点相反；他们的批评者将这场运动斥为“时髦的反垄断”运动。）2021 年，卡恩加入了联邦贸易委员会，预示着当商业模式依赖于收集我们的数据，而产品是免费的时候，反垄断意味着什么将再次受到质疑。
These trends are part of a re-fanging of the state, at least in the United States, as a check on corporate power around data. The broad, multi-sector remit of antitrust has recently been contested, with advocates such as Columbia’s Tim Wu and Lina Khan arguing for an older, “neo-Brandeisian” view that encompasses a much wider understanding of the dangers of economic concentration: monopolies or near monopolies.28 (Louis Brandeis’s view is often opposed to that of Robert Bork; their detractors dismiss this movement as the “hipster antitrust” movement.) In 2021, Khan joined the FTC, presaging a time of renewed contestation of what antitrust means when the business model depends on collecting our data, and the product is free.
“创造互联网的二十六个词”
“THE TWENTY-SIX WORDS THAT CREATED THE INTERNET”
新布兰代斯主义的反垄断法规是可能很快改变国家和企业权力之间平衡的两个主题之一；第二个主题更具体，即要求重新解释“第 230 条”的呼声越来越高。第 230 条指的是 1996 年《通信规范法》中的一句 26 个字的句子：
Neo-Brandeisian antitrust regulation is one of two topics which may soon change the balance between state and corporate power; the second, more data-specific, is growing calls for a reinterpretation of “Section 230.” Section 230 refers to a twenty-six-word sentence within the 1996 Communications Decency Act:
任何交互式计算机服务的提供者或用户不得被视为另一信息内容提供者所提供信息内容的发布者或发言人。
No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider.
这段简短的文字触及了我们上面提到的一个观点：对于一家新企业而言，从行业角度来看，互联网服务提供商 (ISP) 是否被视为内容的“发布者”或“分销商”非常重要。第 230 条是在 Twitter 和 Facebook 等信息平台公司出现之前制定的，这些公司的业务包括对其呈现的内容进行算法排序和优先级排序。然而，其保护措施已被解释为也涵盖这些公司，使它们免于对其通过算法放大、分类和在线分发并从中获利的内容承担法律责任。事实上，第 230 条允许某些形式的企业存在和繁荣。
This brief text touches on a point we made above: for a new business, in a sectoral approach, it matters very much whether, for example, an internet service provider (ISP) is considered a “publisher” or a “distributor” of content. Section 230 was written in a time before information platform companies such as Twitter and Facebook, whose business includes algorithmic sorting and prioritization of the content they present. However, its protections have been interpreted to cover these companies as well, giving them legal immunity from responsibility for the content they algorithmically amplify, sort, and distribute online—and profit from. Indeed, Section 230 has permitted certain forms of businesses to exist and to flourish.
这种豁免是有限的：平台公司仍然会对内容进行审核，包括与恐怖主义相关的内容、露骨的色情内容以及侵犯版权的内容。并非所有审核都是通过算法进行的；人工审核人员（通常称为“内容审核员”而不是编辑）是这一过程的重要组成部分。29纽约大学最近的一份报告估计：
This immunity is finite: platform companies still moderate content, including terrorism-related content, sexually explicit content, and content in violation of copyright. Not all of this moderation is algorithmic; human reviewers, usually termed “content moderators” rather than editors, are a vital part of this process.29 A recent report from NYU estimated:
如今，Facebook 的主平台及其子公司 Instagram 拥有 15,000 名员工，其中绝大多数受雇于第三方供应商。约有 10,000 人负责审查 YouTube 和其他 Google 产品。规模小得多的 Twitter 拥有约 1,500 名版主。30
Today, 15,000 workers, the overwhelming majority of them employed by third-party vendors, police Facebook’s main platform and its Instagram subsidiary. About 10,000 people scrutinize YouTube and other Google products. Twitter, a much smaller company, has about 1,500 moderators.30
在过去的 25 年里，企业和言论自由倡导者都对第 230 条的保护措施表示赞赏。然而，在过去几年里，要求重新评估这一保护措施的呼声在政治左翼和右翼都有所增长。（在撰写本文时，维基百科“第 230 条”页面现在有一个名为“2020 年司法部审查”的部分。）
For the past twenty-five years, corporations and free-speech advocates alike have celebrated the protections of Section 230. In the last few years, however, calls for a reevaluation of this protection have recently grown from both the political left and right. (At the time of writing, the Wikipedia “Section 230” page now has a section called “2020 Department of Justice review.”)
尽管第 230 条适用，但信息平台公司不仅仅是“提供”信息；数据科学家、工程师和产品设计师在此过程中做出的无数主观设计选择正在发挥个性化、优化的编辑功能，即使这些编辑决策是通过算法部署的。没有办法以一种既有用又“中立”的方式呈现如此大量的用户生成内容。随着公民和参议员们对这种算法编辑和放大对社会的影响的认识和担忧日益加深，第 230 条的全面保护范围可能很快导致法院重新解释，甚至出台新的立法。其效果可能是改变这些算法的运作方式内容公司部署了这些技术。具有讽刺意味的是，隐私的使用通常与消费者保护有关，使用端到端加密的信息平台本身免于对内容承担法律责任，因为它们无法通过加密查看内容，因此无法对其进行审核。正如第 230 条“创造了互联网”一样，我们预计这些法律纠纷的解决将对这些公司的运营方式产生巨大影响，从而对它们对社会的影响产生巨大影响。31
Despite Section 230’s applicability, the information platform companies are not merely “providing” the information; the innumerable subjective design choices made by data scientists, engineers, and product designers along the way are performing a personalized, optimized editorial function, even when these editorial decisions are algorithmically deployed. There is no way to present that volume of usergenerated content in a way that is simultaneously useful yet “neutral.” As citizens and senators alike have come to a growing awareness of and concern regarding the impact on society of this algorithmic editing and amplification, the breadth of Section 230’s blanket protections may soon lead to a reinterpretation in the courts or even to novel legislation. The effect could be to change the way these algorithms are deployed by content companies. In an ironic use of privacy, normally associated with consumer protection, information platforms that use end-to-end encryption are themselves protected from legal responsibility for the content, since they are cryptographically unable to view and thus moderate the content. Just as Section 230 “created the internet,” we anticipate that the resolution of these legal contests will have a tremendous effect on the way these companies operate and therefore on their impact on society.31
第三大权力：人民的权力
Power the Third: People Power
我们描述了当前企业权力之间的竞争以及国家权力和数据及数据赋能算法监管的变化性质。公民社会有自己的监管方式。用法律学者的语言来说，这些形成了一种“私人秩序”；我们更宽泛地称之为人民权力。
We’ve described present contests among corporate powers and the changing nature of state power and regulation regarding data and data-empowered algorithms. Civil society has its own ways of regulating. In the language of legal scholars these form a “private ordering”; we’ll refer to these more loosely as people power.
人力：组织内部
PEOPLE POWER: WITHIN ORGANIZATIONS
最明显的“私人订购”形式发生在单个社区（例如公司）内，个人可以产生最直接的影响。在《员工作为监管者：高科技公司的新私人订购》一文中，詹妮弗·范 (Jennifer Fan) 描述了几种可以影响这种订购的机制。32书面倡导就是这样一种机制，现在，通过网络和社交媒体出版的民主化，这种倡导得到了进一步的扩大。这可以包括直接与媒体沟通的作用，通常速度之快可以“实时”产生影响。 《大西洋月刊》引用了这样一个例子，作为泄密时间表的一部分：
The most visible form of “private ordering,” where individual people can have the most direct impact, occurs within single communities, such as a company. In “Employees as Regulators: The New Private Ordering in High Technology Companies,” Jennifer Fan describes several mechanisms that can effect this ordering.32 Written advocacy, now amplified by the democratization of publishing via the web and social media, is one such mechanism. This can include the role of communicating directly with the press, often at a speed which makes such an impact “real time.” One such example is quoted in The Atlantic, as part of a timeline of leaks:
2018 年 8 月 17 日，谷歌致纽约时报：谷歌的一名员工引用了Sundar Pichai 在一次演讲中对纽约时报记者Kate Conger 的言论关于 Dragonfly 项目。Conger 发推文，导致一名 Google 员工在公开场合说了 [脏话]，这也被泄露了。33
August 17, 2018, Google to The New York Times: Some Google employee feeds The Times’ Kate Conger lines from a talk that Sundar Pichai was giving about the Dragonfly project. Conger tweets them, leading one Googler to say [an expletive] on the open mic, which was also leaked.33
这种宣传作为集体行动的一部分尤其有用，例如最近美国有线电视新闻网（CNN）采访了《纽约时报》的一位数据分析师，解释了员工对成立工会的兴趣。
Such publicity can be particularly useful as part of collective action, for example a recent CNN interview with a data analyst at The New York Times, explaining the employees’ interest in unionizing.
另一种需要集体行动的私人命令形式是收集同事的私人信息。工资信息是希望平衡员工激励与薪酬公平的公司所秘密掌握的信息，可以成为集体行动的一个特别有力的工具。这种做法暴露了不同人口群体之间的不平等，谷歌的 Erica Baker 在 2015 年开始收集工资数据电子表格时就是这种情况。34
Another form of private ordering requiring collective action involves gathering private information from colleagues. Salary information, closely held information by companies wishing to balance employee incentives with pay equity, can be a particularly powerful tool for collective action. Such practices expose inequality across demographic groups, as was the case when Erica Baker of Google began collecting a spreadsheet of salary data in 2015.34
从更大范围来看，在上市公司中，股东积极行动提供了另一种私人秩序机制。在许多公司中，员工也是股东，这引发了一些强有力的时刻，例如亚马逊股东大会上员工积极行动，讨论各种道德问题。话虽如此，科技公司越来越多地采用较旧的两层股票模式，其中包括《纽约时报》，其中某些股票比其他股票拥有更多的投票权。尽管这些公司是“上市公司”，但这种制度让 Facebook、亚马逊和 Snapchat 等公司的创始人对其公司拥有超大控制权。
At a larger scale, in publicly traded companies, shareholder activism offers another mechanism of private ordering. In many companies, employees are also shareholders, which has given rise to powerful moments such as employee activism at Amazon shareholder meetings around a variety of ethical issues. That said, technology companies have increasingly adopted the older, two-tiered stock model of, among others, The New York Times, with certain shares having more voting rights than others. This system gives the founders of, for example, Facebook, Amazon, and Snapchat outsize control of their companies despite the companies being “public.”
跨公司联盟，例如技术工人联盟（成立于 2014 年）、技术团结（2016 年）、Never-again.tech（2016 年）和监狱技术抵抗网络（CTRN，2020 年），正在教育和组织技术工人，要求相关公司进行变革。促进集体行动的非营利组织，例如 Coworker.org，允许这些公司的员工向管理层施加压力。随着对工程人才的需求越来越大，公司中，招聘和留住这些员工的困难威胁正日益被利用。在《数据女权主义》一书中，迪格纳齐奥和克莱因指出：“数据人一般都有选择——选择为谁工作、从事哪些项目以及拒绝哪些价值观。” 35与许多员工相比，他们更能坚持要求改变数据未来。对于那些尚未准备好离开雇主的人来说，罢工是一种集体行动，谷歌就是一个著名的例子，2018 年 11 月 1 日上午，超过 20,000 名员工举行了一场引人注目的抗议活动。*然而，这些罢工最终并没有带来组织者所倡导的改变，两位主要组织者（克莱尔·斯台普尔顿和梅雷迪斯·惠特克）以谷歌的报复为由辞职。尽管如此，他们仍然坚持，公开发表声明说明他们的选择，并与媒体讨论他们的抱怨。
Cross-company coalitions such as the Tech Workers Coalition (founded in 2014), Tech Solidarity (2016), Never-again.tech (2016), and the Carceral Tech Resistance Network (CTRN, 2020) are educating and organizing tech workers to demand change across related companies. Nonprofits facilitating collective action, such as Coworker.org, allow employees of such companies to put pressure on management. As the demand for engineering talent puts increasing strain on companies, the threat of difficulty in recruiting and retaining these workers is increasingly being leveraged. In Data Feminism, D’Ignazio and Klein note, “Data people, generally speaking, have choices—choices in who they work for, which projects they work on, and what values they reject.”35 More than many employees, they can insist on alternate data futures. One form of collective action for those not ready to leave their employers is the walkout, with notable examples at Google, where more than 20,000 employees staged a highly visible protest on the morning of November 1, 2018.* Ultimately, however, these walkouts did not bring about the changes the organizers were advocating, and two of the main organizers (Claire Stapleton and Meredith Whittaker) resigned, citing retaliation by Google. Nevertheless, they persisted, publishing public statements about their choices, and discussing with the press their complaints.
目前，科技员工越来越多地转向集体行动，尤其是工会化。这包括 2020 年 Kickstarter 和 Glitch 的工会化，以及 2021 年 Alphabet、亚马逊、道琼斯和《纽约时报》等公司的科技员工的工会化努力。
At present, tech employees increasingly are turning not just to collective action but specifically toward unionization. This includes 2020 unionization at Kickstarter and Glitch, and 2021 unionization efforts by tech workers at, among other companies, Alphabet, Amazon, Dow Jones, and The New York Times.
虽然个人或集体采取的此类行动不会像 GDPR 那样产生全面的监管变化的直接影响，但它们可能会产生巨大的影响，特别是对于软件工程师等关键团队中有很大一部分参与的公司。正如 CTRN 的 Sarah T. Hamid 所写，“我们正在对抗的系统已经存在很长时间了。。。但如果你能引入一点摩擦，你可以腾出一些喘息的空间。” 36鲁哈·本杰明 (Ruha Benjamin)在她的著作《科技之后的竞赛》中反对轻易的技术修复，她指出，“我们必须要求技术设计师和决策者成为负责任的技术管理者，能够促进社会福利”，她以算法正义联盟的安全面容承诺为例。37这样的努力可以在内部进行，也可以在外部进行。
While such actions, by individuals or collectively, do not have the immediate impact of sweeping regulatory changes like GDPR, they can have tremendous effects, particularly for companies in which a large fraction of a critical team, such as software engineers, participate. As Sarah T. Hamid of CTRN writes, “The systems we’re fighting have been around for a long time. . . . But if you can introduce a bit of friction, you can open up some breathing room.”36 In her Race After Technology, Ruha Benjamin, pushing against facile tech fixes, notes, “we must demand that tech designers and decision-makers become accountable stewards of technology, able to advance social welfare” using the example of the Safe Face Pledge of the Algorithmic Justice League.37 Such efforts can happen internally—and externally.
人力：外部
PEOPLE POWER: EXTERNAL
现在，我们公众应该认真对待我们对人工智能当前和迫在眉睫的社会后果所负的责任。……责任不能简单地被推卸或外包给技术开发商和私营公司。……公民必须将围绕人工智能的问题视为我们所有人的集体问题，而不仅仅是他们（公司和国家）的技术问题。
It is high time for us as a public to take seriously our responsibilities for the present and looming social consequences of AI. . . . Responsibility cannot simply be offloaded and outsourced to tech developers and private corporations. . . . Citizens must come to view issues surrounding AI as a collective problem for all of us rather than a technical problem just for them [corporations and the state].
—Annette Zimmermann、Elena Di Rosa 和 Hochan Kim 38
—Annette Zimmermann, Elena Di Rosa, and Hochan Kim38
广大公众也与这些公司打交道，他们既是用户，又是以训练数据形式提供的免费劳动力，更广泛地说，他们也是我们宝贵的行为和个人数据的供应商。对于 Spotify 或 Netflix 等公司，公众通过订阅直接为公司提供资金。对于以我们为产品而非客户的公司，我们享受服务带来的便利，但我们很少“问自己作为公众在授权和反对公司和国家使用人工智能技术方面的角色问题”，齐默曼、迪罗莎和金指出。39个人的外部行动，即使是广泛的，也很少对公司产生明显的影响。据 #DeleteUber 报道，2017 年 1 月的 #DeleteUber 运动导致“数十万”用户删除他们的应用程序该公司的 IPO 申请文件。事实上，这发生在公司糟糕的一年之初，包括创始人兼首席执行官被替换，尽管这只是公司面临的几个麻烦之一。这样的公开行动不仅剥夺了公司的收入，也剥夺了宝贵的数据——“数据抵制”；人才的辞职则带来了“人才抵制”。用哈米德的话来说，这些都带来了少量的“摩擦”。数据未来的一个关键问题是，足够的集体内部或外部摩擦是否会减缓数据驱动型技术公司影响力和实力的增长。
The wider public engage with these companies as well, as users, as providers of free labor in the form of training data, and as suppliers of our valuable behavioral and personal data more generally. For companies such as Spotify or Netflix, the public supplies the company’s funding directly through subscriptions. For companies for which we are the product rather than the customer, we enjoy the convenience of the service, but we too rarely “ask uncomfortable questions about our own role as a public,” Zimmermann, Di Rosa, and Kim point out, “in authorizing and contesting the use of AI technologies by corporations and the state.”39 External action by individuals, even when widespread, rarely has a visible impact on a company. The #DeleteUber movement in January 2017 led to “hundreds of thousands” of users deleting their app, according to the company’s IPO filing documents. Indeed, this took place at the start of a terrible year for the company, including the founder-CEO being replaced, though this was one of only several troubles for the company. Such a public action denies a company revenue, but also valuable data—a “data boycott”; the resignations of talent provide a “talent boycott.” Each of these provides a small amount of “friction,” to use Hamid’s language. A pivotal question in the future of data will be whether sufficient collective internal or external friction serves to slow the growth in reach and power of data-driven technology companies.
这些公司带来的最严重危险可能是它们破坏了民主进程本身的可能性。耶鲁大学法学教授艾米·卡普钦斯基写道：“当今的信息资本主义不仅威胁到我们的个人主观性，也威胁到平等和自治能力。数据和民主的问题，而不仅仅是数据和尊严的问题，必须成为我们关注的核心。”40 私人秩序，包括集体行动，有时是愤世嫉俗的联盟，在各个治理层面，拥有各种形式的权力，可能成为一条道路，让我们将算法系统转向加强自治，以实现正义，而不是进一步破坏我们的治理并加剧现有的不平等。
The gravest dangers from those companies may be that they undermine the possibility of the democratic process itself. “Today’s informational capitalism,” Yale law professor Amy Kapczynski writes, “brings a threat not merely to our individual subjectivities but to equality and our ability to self-govern. Questions of data and democracy, not just data and dignity, must be at the core of our concern.”40 Private ordering, including collective action, at times cynical coalitions, at all levels of governance, with all the sundry forms of power, might become the path that will allow us to turn algorithmic systems toward enhancing self-governance for the purpose of achieving justice, rather than destroying our governance further and heightening existing inequalities.
重返不稳定的游戏
Return to the Unstable Game
永远不要预言，尤其是关于未来的预言。
Never prophesy, especially about the future.
— 归功于塞缪尔·戈德温和/或尼尔斯·玻尔
—Attributed to Samuel Goldwyn and/or Niels Bohr
我们并没有试图预言或提倡革命，而是关注权力之间的竞争——企业权力、国家权力和人民权力。这些权力之间的权力在时间和效力的尺度上发挥着非常不同的作用，尽管每一个都有可能塑造数据的未来。数据仍然是当权者（尤其是国家和企业）控制其领域的极其有效的方式。就数据支持技术而言，我们可能很难记起以前的时代——智能手机出现之前、万维网出现之前、家中出现 24/7 监控设备之前的时代。历史观让现在变得奇怪，因为它打破了技术决定论的谬论：即技术导致社会、经济和文化变革的信念。要使技术发挥其影响，就需要法律、基础设施和社会决策，使技术能够成长并成为我们规范的一部分。这些影响不会凭空而来。
Instead of attempting prophecy or advocating revolution, we’ve focused on contests among powers—corporate power, state power, and people power. The forces among these powers act on very different scales of time and efficacy, though each has potential to shape the future of data. Data continues to be a tremendously effective way for incumbent powers to maintain control over their domains—particularly state and corporate incumbents. In the case of data-enabled technologies, it can be difficult for us to remember a time before— a time before smartphones, before the WWW, before 24/7 surveillance devices in our homes. A historical view makes the present strange, as it shatters the fallacy of technological determinism: the belief that technology causes social, economic, and cultural transformation. For technologies to have the impact they have, they require legal, infrastructural, and social decision-making that enable technologies to grow and become part of our norms. Those effects don’t just happen.
我们希望，这里呈现的潜在未来观（一种不稳定的游戏）能提醒你，现在不是监禁，而只是我们当前的快照：我们不必使用不道德或不透明的算法决策系统，即使在技术上可行的情况下也是如此。基于大规模监视的广告不是我们社会的必要元素。我们不需要建立学习过去和现在的分层并在未来强化它们的系统。隐私不会因为技术而消亡；支持新闻、书籍写作或任何对你来说重要的手艺的唯一方法并不是监视你以投放广告。还有其他选择。这些系统中的许多都包含了我们社会想要的元素，但其中许多并没有。这需要时间。工作将是细致入微的。这些事情都不是一蹴而就的。它不会像在成本函数中添加一个新术语那么简单。它不会是一项监管裁决。这可能会涉及奇怪的、有时是愤世嫉俗的、甚至不舒服的联盟。新兴技术通常首先提供给掌权者；有时他们会利用这些技术来帮助被压迫和被剥夺权利的人，但通常他们新兴技术可以增强弱势群体的自信心，并利用这些技术来捍卫和扩大自己的权力和控制力。因此，需要一段时间才能重新调整规范、法律、架构和市场，以利用这些新兴能力来赋予弱势群体权力，但这是可以做到的。41技术意味着变革，但社会变革需要时间：正如我们所见，一项技术有时需要几十年才能融入社会，才能符合我们的价值观和规范，甚至完全符合。我们可以直接或间接地利用许多潜在权力，无论大小，来塑造技术与规范、法律与市场以及数据在其中的作用之间的关系。
We hope that the view of the potential future presented here, as an unstable game, reminds you that the present is not a prison sentence, but is merely our current snapshot: We don’t have to use unethical or opaque algorithmic decision systems, even in contexts where their use may be technically feasible. Ads based on mass surveillance are not necessary elements of our society. We don’t need to build systems that learn the stratifications of the past and present and reinforce them in the future. Privacy is not dead because of technology; it’s not true that the only way to support journalism or book writing or any craft that matters to you is spying on you to serve ads. There are alternatives. Many of these systems incorporate elements that we want in our society, and many of them do not. It will take time. The work will be granular. None of these things are going to be quick fixes. It’s not going to be as simple as adding a new term to a cost function. It’s not going to be one regulatory ruling. It’s likely to involve strange and sometimes cynical, even uncomfortable alliances. Emerging technologies are generally first available to those who are in power; sometimes they use these to enable the oppressed and disenfranchised, but often they use them to defend and extend their own power and control. So it takes a while to reorient norms, laws, architecture and markets in a way that harnesses these emerging capabilities in order to empower the defenseless—but it can be done.41 Technology means change, but societal change takes time: as we’ve seen, sometimes it takes decades for a technology to get integrated into society before it comports with our values and norms—if it does at all. Many potential forces, large and small, are available to us, directly and indirectly, to shape the relationships among technology and norms, laws and markets, and data’s role in it all.
*有关 2016 年至 2018 年期间此类私人订购的例子的时间线，以及对九名科技员工的采访，包括来自科技工人联盟和 2018 年谷歌罢工的员工，请参阅 Cameron Bird 等人撰写的《科技起义》，《加州周日杂志》 ，2019 年 1 月 23 日， https://story.californiasunday.com/tech-revolt/。
* For a timeline of examples of such private ordering from 2016 to 2018, along with interviews with nine tech employees, including from Tech Workers Coalition and the 2018 Google walkout, see Cameron Bird et al., “The Tech Revolt,” California Sunday Magazine, January 23, 2019, https://story.californiasunday.com/tech-revolt/.
这本书源自我们为哥伦比亚大学和巴纳德学院本科生开设的一门课程。我们最初设想在哥伦比亚大学学生的倡议下教授这门课程，并得到哥伦比亚大学合作实验室的财政、道德和行政支持，该实验室由理查德·威滕和时任哥伦比亚大学数据科学研究所所长的珍妮特·温领导。如果不是我们 2017 年至 2022 年的学生，这项工作的范围会更加狭窄和有限。他们不断推动我们完善对数据、真相和权力之间持续冲突的描述，从十八世纪到今天。我们感谢他们的专注、参与和好奇心。他们的问题塑造了材料，并激励我们更加努力地寻找历史和技术根源，这些根源有助于“解释”现在，但通过描绘本来很容易实现的反事实现实，使现在变得奇怪。这些原本可能存在的其他世界也激发了我们对我们所有人可能创造和享受的未来的讨论。
This book emerged from a class we developed for Columbia and Barnard undergraduates. We first envisioned teaching it at the instigation of a group of Columbia students, and with the financial, moral, and bureaucratic support provided by the Collaboratory at Columbia, led by Richard Witten and Jeanette Wing, then director of Columbia’s Data Science Institute. Were it not for our students from 2017 to 2022, this work would have been far more narrow and limited in scope. They consistently pushed us to sharpen our account of the persistent conflicts among data, truth, and power, from the eighteenth century to present day. We’re grateful for their focus, engagement, and curiosity. Their questions shaped the material and inspired us to work harder to find the historical and technological roots that help “explain” the present yet make the present strange by picturing the counterfactual realities that easily could have been. These other worlds, those that could have been, also inspired our discussions of the possible futures we all may yet create and enjoy.
我（威金斯）对在工业中开发和部署机器学习的现实情况的理解，受到了《纽约时报》数据科学团队过去和现在的优秀同事的影响，以及各个信息平台公司中同样熟练的技术人员，他们可能更希望我不说出他们的名字，但知道他们是谁。
MY (WIGGINS’S) UNDERSTANDING of the realities of developing and deploying machine learning in industry has been shaped by my fantastic colleagues, past and present, on the Data Science team at The New York Times, along with similarly skilled technologists at various information platform companies who would probably prefer I not name them but know who they are.
我对道德的理解最初源于与 Matt Salganik 就他的书《Bit by Bit》的讨论，后来通过与《纽约时报》数据治理主管 Robin Berjon 的多次讨论而不断完善，尤其是在 Salganik 教授担任《纽约时报》驻校学者的那一年。我感谢 David Blei、David Donoho、Gerd Gigerenzer、Mark Hansen、Gina Neff、Peter Norvig、Cathy O’Neil、Deb Raji、Ben Recht、Alfred Spector、Latanya Sweeney、Anne Washington、Hadley Wickham 和 Jeanette Wing 就数据、数据科学和道德提出的额外见解和推动性问题。多年来与 David Carroll、Renee DiResta、Joan Donovan 和 Justin Hendrix 的交谈有助于理解算法如何塑造和扭曲现实。与 Mark Thompson 的多次交谈极大地澄清了数据对我们集体判断真相的更广泛影响，尤其是在媒体和政治领域。
My understanding of ethics was initially forged from discussions with Matt Salganik about his book Bit by Bit, and refined by many discussions with Robin Berjon, director of data governance at The Times, particularly during Professor Salganik’s year as scholar in residence at The Times. I thank David Blei, David Donoho, Gerd Gigerenzer, Mark Hansen, Gina Neff, Peter Norvig, Cathy O’Neil, Deb Raji, Ben Recht, Alfred Spector, Latanya Sweeney, Anne Washington, Hadley Wickham, and Jeanette Wing for additional insights and driving questions on data, data science, and ethics. Understanding the ways algorithms shape and distort reality was helped by conversations for years with David Carroll, Renee DiResta, Joan Donovan, and Justin Hendrix. The broader impact of data on our collective truthmaking, particularly in the arena of media and politics, was greatly clarified by many conversations with Mark Thompson.
我感谢 Matt Jones、Ariel Kaminer、Rob Phillips 和 Allison Schrager 就如何写书给出的无尽建议。当然，我还要永远感谢我的父母 Richard 和 Carolyn Wiggins。
I thank Matt Jones, Ariel Kaminer, Rob Phillips, and Allison Schrager for boundless advice on how actually to write a book. And thank you of course and always to my parents, Richard and Carolyn Wiggins.
我们一直试图向那些改变了我们思维的学者、政策制定者和技术人员表达我们的感激之情。我（琼斯）非常幸运地与斯蒂芬妮·迪克、理查德·斯塔利、穆斯塔法·阿里、乔尼·佩恩和莎拉·狄龙合作完成了为期两年的项目《人工智能史：权力谱系》，这是梅隆-索耶研讨会，在此之前，我与迪克、佩恩和亚伦·门登-普拉塞克合作举办了一场关于人工智能历史的研讨会。在整个过程中，来自世界各地的鼓舞人心的学者对我们提出了挑战，也给了我们启发。我受益匪浅来自 Data & Society 的奖学金，在那里我有幸与 Danah Boyd、Darakhshan Mir、Jeanna Matthews、Seth Young 和 Claudia Haupt 合作。除了出色的奖学金之外，Seth Young 还在思考问责制方面提供了许多指导——与他合作的一个未完成的项目使这里的其中一章成为可能。在哥伦比亚大学，Eben Moglen 好心地允许我参加两门法律课程，Rachel Schutt 和 Cathy O'Neil 允许我参加他们的第一门数据科学课程。Chris Wiggins 参加了我关于这些主题的第一次初步演讲——我在那里第一次遇到的思路清晰和慷慨贯穿了我们整个合作过程，最终在这本书中达到顶峰。
WE’VE TRIED THROUGHOUT to express our debts to scholars, policymakers, and technologists who have transformed our thinking. I (Jones) have been exceptionally lucky to work with Stephanie Dick, Richard Staley, Mustafa Ali, Jonnie Penn, and Sarah Dillon on the two-year project History of Artificial Intelligence: A Genealogy of Power, a Mellon-Sawyer seminar, and before that with Dick, Penn, and Aaron Mendon-Plasek on a workshop on the history of AI. Throughout the process, inspiring scholars from around the world challenged and inspired us. I benefited greatly from a fellowship at Data & Society, where I had the great fortune to work especially with danah boyd, Darakhshan Mir, Jeanna Matthews, Seth Young, and Claudia Haupt. In addition to great fellowship, Seth Young provided much guidance in thinking about accountability—an unfinished project with him made one of the chapters here possible. At Columbia, Eben Moglen kindly allowed me to take two law courses, and Rachel Schutt and Cathy O’Neil permitted me to attend their first data science course. Chris Wiggins attended my first preliminary talk on these subjects—the clarity of thought and generosity I first encountered there has run through our entire collaboration, culminating in this book.
这里介绍的数据科学史的关键方面汇集在柏林马克斯普朗克科学史研究所和加州大学洛杉矶分校在亨廷顿举办的研讨会上。这里对思想的更多学术迭代受益于来自密歇根大学安娜堡分校、印第安纳大学、加州大学伯克利分校、加州大学洛杉矶分校、加州大学圣塔芭芭拉分校、加州大学圣地亚哥分校、南洋理工大学、巴黎政治学院、康奈尔科技大学、剑桥大学、欧洲大学研究所、芝加哥大学、锡根大学、宾夕法尼亚大学、罗格斯大学和哥伦比亚大学计算机科学系的明智听众的提问和评论。我从 David Isenberg 周围的社区受益匪浅。我的研究生 Aaron Mendon-Plasek 一直深入研究机器学习的历史；他的工作很快就会超越这里给出的叙述。
Key facets of the history of the data sciences presented here came together for workshops at the Max Planck Institute for History of Science in Berlin and at the Huntington hosted by UCLA. More academic iterations of ideas here have benefited enormously from questions and comments from sage audiences at University of Michigan–Ann Arbor; Indiana University; UC Berkeley, UCLA, UCSB, UCSD; Nanyang Technological University, Sciences Po; Cornell Tech; University of Cambridge; the European University Institute; University of Chicago; University of Siegen; University of Pennsylvania; Rutgers; and the Department of Computer Science at Columbia. I’ve greatly benefited from communities around David Isenberg. My graduate student Aaron Mendon-Plasek has been working deep in the history of machine learning; his work will soon eclipse the accounts given here.
承担这个项目意味着重返校园，梅隆基金会和古根海姆基金会为这个项目提供了支持。斯隆基金会资助了哥伦比亚大学举办的一系列关于数据和人工智能历史的研讨会。
Undertaking this project meant returning to school, something made possible by Mellon and Guggenheim foundations. The Sloan Foundation funded a series of workshops at Columbia on histories of data and of artificial intelligence.
斯坦福特别馆藏中心、大英图书馆、美国哲学学会、哥伦比亚特别馆藏中心、马萨诸塞大学阿默斯特分校、普林斯顿大学贝克图书馆和内华达大学里诺特别馆藏中心的档案管理员，使得这项研究成为可能。许多政府雇员回答了《信息自由法》的要求，同样确保了这里的大部分历史可以得到讲述。
Archivists at Stanford Special Collections, the British Library, the American Philosophical Society, Columbia Special Collections, UMass-Amherst, the Baker Library at Princeton, and the University of Nevada, Reno Special Collections, made the research possible. The many government employees who answer FOIA requests likewise ensured that much of the history here could be told.
我的三个女儿热爱读书，也许有一天她们会读到这本书——而且毫无疑问，她们会发现书中仍然存在的错误和不妥之处。伊丽莎白·李的洞察力、爱和智慧贯穿了我所写和所做的一切。
With their intense love of books, my three daughters may read this one someday—and, without doubt, uncover the remaining errors and infelicities. Elizabeth Lee’s insight, love, and intelligence echo through everything I write and do.
我们非常感谢 Ella Coon 和 Susannah Glickman，他们在校对、编辑和改进本书的最终草稿方面付出了超出职责范围的努力。慷慨的同事 Stephanie Dick、danah boyd、Theodore Porter、David Sepkoski 和 Sarah Igo 仔细阅读并大大改进了草稿章节，指出了我们的错误并鼓励了我们的工作。Chris Eoyang、Su Hang、Willian Janeway、DJ Patil 和 JB Rubinovitz 也对本书的草稿提供了出色的批评意见。其余所有错误都是我们的错。斯隆基金会支持了本书的完成；特别感谢 Josh Greenberg 的鼓励。
WE BOTH THANK Ella Coon and Susannah Glickman, who went far beyond the call of duty in correcting, editing, and improving the final draft of the book. Generous colleagues Stephanie Dick, danah boyd, Theodore Porter, David Sepkoski, and Sarah Igo carefully read and greatly improved draft chapters, noting our errors and encouraging the work. Excellent critical comments on drafts of this book were likewise provided by Chris Eoyang, Su Hang, Willian Janeway, DJ Patil, and JB Rubinovitz. All remaining errors are entirely our fault. The Sloan Foundation supported the completion of this book; thanks especially to Josh Greenberg for his encouragement.
2014 年，Cathy O'Neil 和 Mark Hansen 邀请我们作为哥伦比亚新闻学院 Lede 项目的一部分，为数据记者授课，从而将关键数据科学付诸实践。这是我们第一次有机会共同开发课程并共同教学；课程结构（将讲座和功能性互动与 Python 材料相结合）以及关于数据在社会和真相探求中的作用的问题都为我们后来的课程提供了参考。
Cathy O’Neil and Mark Hansen invited us in 2014 to teach data journalists as part of Columbia Journalism School’s Lede Program, and thus to put into practice critical data science. This was our first chance to develop a curriculum together and to teach together; both the structure— mixing lecture and functional engagement with material in Python—and the questions asked about the role of data in society and in truthmaking informed our later course.
我们的经纪人 Eric Lupfer 将我们最初的离心想法塑造成一份提案，然后将其塑造成一本具有强烈叙述和论证重点的书。我们的编辑 John Glusman 凭借他的洞察力和精准度，帮助将初稿变成了您面前的这本书。
Our agent Eric Lupfer shaped our initial centrifugal ideas into a proposal and then a book with a strong narrative and argumentative focus. With his insight and precision, our editor John Glusman helped turn a first draft into the book before you.
1. Kevin Roose 和 Cecilia Kang，“马克·扎克伯格在持怀疑态度的立法者面前就 Facebook 问题作证”，《纽约时报》，2018 年 4 月 11 日，第 US 版，https://www.nytimes.com/2018/04/10/us/politics/zuckerberg-facebook-senate-hearing.html。
1. Kevin Roose and Cecilia Kang, “Mark Zuckerberg Testifies on Facebook Before Skeptical Lawmakers,” New York Times, April 11, 2018, sec. US, https://www.nytimes.com/2018/04/10/us/politics/zuckerberg-facebook-senate-hearing.html.
2.课程每周安排两次单独的会议，周二进行讨论，周四进行实用材料操作：即使用 Python 进行计算操作，执行周二讨论的数据分析和机器学习模型。我们并没有试图在本书中捕捉课程的应用部分，但我们邀请有兴趣仔细阅读材料的读者直接使用课程网站上提供的数据和代码：https://data-ppf.github.io/。
2. The weekly cadence of the class involved two separate meetings, with discussions on Tuesdays and functional engagement with the material on Thursdays: that is, computational engagement in Python performing the types of data analyses and machine learning models discussed on Tuesdays. We’ve not attempted in this book to capture the applied half of the class, but we invite readers who are interested in a closer read of the material to engage directly with the data and the code provided online at the course website: https://data-ppf.github.io/.
3.我们这里的语言借用了 Phillip Rogaway 的《密码工作的道德品质》（2015），1，https://web.cs.ucdavis.edu/~rogaway/papers /moral-fn.pdf。
3. Our language here borrows from Phillip Rogaway, “The Moral Character of Cryptographic Work” (2015), 1, https://web.cs.ucdavis.edu/~rogaway/papers /moral-fn.pdf.
第一章：利害关系
CHAPTER 1: THE STAKES
1. Hanna Wallach，“大数据、机器学习和社会科学”，Medium，2014 年 12 月 23 日，https://medium.com/@hannawallach/big-data-machine-learning-and-the-social-sciences-927a8e20460d。
1. Hanna Wallach, “Big Data, Machine Learning, and the Social Sciences,” Medium, December 23, 2014, https://medium.com/@hannawallach/big -data-machine-learning-and-the-social-sciences-927a8e20460d.
2.瓦拉赫。
2. Wallach.
3. danah boyd 和 Kate Crawford，《大数据的关键问题》，《信息、通信与社会》第 15 卷，第 5 期（2012 年 6 月 1 日）：663，https://doi.org/10.1080/1369118X.2012.678878。
3. danah boyd and Kate Crawford, “Critical Questions for Big Data,” Information, Communication & Society 15, no. 5 (June 1, 2012): 663, https://doi .org/10.1080/1369118X.2012.678878.
4.这种倡导呼应了 20 世纪 60 年代工程师们寻求更具社会和环境意识的技术的运动，详见 Matthew H. Wisnioski 所著的《变革的工程师：20 世纪 60 年代美国的技术竞争愿景》（马萨诸塞州剑桥：麻省理工学院出版社，2012 年）。
4. Such advocacy echoes movements in the 1960s among engineers seeking more social and environmentally conscious technologies, chronicled in Matthew H. Wisnioski, Engineers for Change: Competing Visions of Technology in 1960s America (Cambridge, MA: MIT Press, 2012).
5. Safiya Umoja Noble，《谷歌搜索：超高可见度是让黑人妇女和女孩隐形的一种手段》，《InVisible Culture》 ，第 19 期（2013 年 10 月 29 日），http://ivc.lib.rochester.edu/google-search-hyper-visibility-as-a-means-of-rendering-black-women-and-girls-invisible/。她在《压迫算法：搜索引擎如何强化种族主义》（纽约：纽约大学出版社，2018 年）中阐述了这些论点。
5. Safiya Umoja Noble, “Google Search: Hyper-Visibility as a Means of Rendering Black Women and Girls Invisible,” InVisible Culture, no. 19 (October 29, 2013), http://ivc.lib.rochester.edu/google-search-hyper-visibility-as-a -means-of-rendering-black-women-and-girls-invisible/. She developed the arguments in her Algorithms of Oppression: How Search Engines Reinforce Racism (New York: New York University Press, 2018).
6. Cathy O'Neil，《数学毁灭性武器：大数据如何加剧不平等并威胁民主》（纽约：Crown，2016 年），第 48 页。
6. Cathy O’Neil, Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy (New York: Crown, 2016), 48.
7. Ruha Benjamin，《科技之后的种族：新吉姆法典的废奴主义工具》（英国剑桥；马萨诸塞州梅德福：Polity Press，2019 年），第 44–45 页。
7. Ruha Benjamin, Race after Technology: Abolitionist Tools for the New Jim Code (Cambridge, UK; Medford, MA: Polity Press, 2019), 44–45.
8. Meredith Whittaker，“捕获的高昂成本”，Interactions 28，第 6 期（2021 年 11 月）：50–55，https://doi.org/10.1145/3488666。
8. Meredith Whittaker, “The Steep Cost of Capture,” Interactions 28, no. 6 (November 2021): 50–55, https://doi.org/10.1145/3488666.
9. Virginia Eubanks，《公共思想家：Virginia Eubanks 论数字监控和人民权力》，采访人：Jenn Stroud Rossman，公共图书（在线），2020 年 7 月 9 日，https://www.publicbooks.org/public-thinker-virginia-eubanks-on-digital-surveillance-and-people-power/。
9. Virginia Eubanks, “Public Thinker: Virginia Eubanks on Digital Surveillance and People Power,” interview by Jenn Stroud Rossman, Public Books (online), July 9, 2020, https://www.publicbooks.org/public-thinker -virginia-eubanks-on-digital-surveillance-and-people-power/.
10. Lisa Nakamura，《互联网是一场垃圾大火。以下是解决方法》，2019 年，https://www.ted.com/talks/lisa_nakamura_the_internet_is_a_trash_fire _here_s_how_to_fix_it。
10. Lisa Nakamura, The Internet Is a Trash Fire. Here’s How to Fix It, 2019, https://www.ted.com/talks/lisa_nakamura_the_internet_is_a_trash_fire _here_s_how_to_fix_it.
11. Zeynep Tufekci，“工程化公众：大数据、监控和计算政治”，《第一个星期一》，2014 年 7 月 2 日，https://doi.org/10.5210/fm.v19i7.4901。
11. Zeynep Tufekci, “Engineering the Public: Big Data, Surveillance and Computational Politics,” First Monday, July 2, 2014, https://doi.org/10.5210/fm .v19i7.4901.
12. Renee DiResta，“调解同意”，ribbonfarm（博客），2019 年 12 月 17 日，https://www.ribbonfarm.com/2019/12/17/mediating-consent/。
12. Renee DiResta, “Mediating Consent,” ribbonfarm (blog), December 17, 2019, https://www.ribbonfarm.com/2019/12/17/mediating-consent/.
13. Virginia Eubanks，《自动化不平等：高科技工具如何剖析、监管和惩罚穷人》（纽约：圣马丁出版社，2017 年）。
13. Virginia Eubanks, Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor (New York: St. Martin’s Press, 2017).
14. Brianna Posadas，“芝加哥的‘战略主题名单’有多具战略性？Upturn 调查”，Medium，2017 年 6 月 26 日，https://medium.com/equal-future/how-strategic-is-chicagos-strategic-subjects-list-upturn-investigates-9e5b4b235a7c。
14. Brianna Posadas, “How Strategic Is Chicago’s ‘Strategic Subjects List’? Upturn Investigates,” Medium, June 26, 2017, https://medium.com/equal -future/how-strategic-is-chicagos-strategic-subjects-list-upturn-investigates -9e5b4b235a7c.
15. 请参阅 Martha Poon，“企业资本主义与大数据日益增长的权力：评论文章”，《科学、技术与人类价值观》第 41 卷第 6 期（2016 年）：1088–1108 页。
15. See Martha Poon, “Corporate Capitalism and the Growing Power of Big Data: Review Essay,” Science, Technology, & Human Values 41, no. 6 (2016): 1088–1108.
16. Whittaker，“捕获的高昂成本”；Rodrigo Ochigame，“‘道德人工智能’的发明：大型科技公司如何操纵学术界以避免监管”，The Intercept（博客），2019 年 12 月 20 日，https://theintercept .com/2019/12/20/mit-ethical-ai-artificial-intelligence/；Thao Phan 等人，“美德经济：大型科技公司中‘道德’的流通”，科学作为文化，2021 年 11 月 4 日，第 1-15 页，https://doi.org/10.1080/09505431.2021 .1990875； Matthew Le Bui 和 Safiya Umoja Noble，《我们缺少人工智能的正义道德框架》，《牛津人工智能伦理手册》 ，Markus Dirk Dubber、Frank Pasquale、Sunit Das 编辑。（牛津：牛津大学出版社，2020 年），https://doi.org/10.1093/oxfordhb /9780190067397.013.9。
16. Whittaker, “The Steep Cost of Capture”; Rodrigo Ochigame, “The Invention of ‘Ethical AI’: How Big Tech Manipulates Academia to Avoid Regulation,” The Intercept (blog), December 20, 2019, https://theintercept .com/2019/12/20/mit-ethical-ai-artificial-intelligence/; Thao Phan et al., “Economies of Virtue: The Circulation of ‘Ethics’ in Big Tech,” Science as Culture, November 4, 2021, 1–15, https://doi.org/10.1080/09505431.2021 .1990875; Matthew Le Bui and Safiya Umoja Noble, “We’re Missing a Moral Framework of Justice in Artificial Intelligence,” The Oxford Handbook of Ethics of AI, Markus Dirk Dubber, Frank Pasquale, Sunit Das, eds. (Oxford: Oxford University Press, 2020), https://doi.org/10.1093/oxfordhb /9780190067397.013.9.
17. 有关私人秩序，请参阅 Jennifer S Fan，《员工作为监管者：高科技公司的新私人秩序》，《犹他法律评论》 ，第 5 期（2019 年）：55。
17. For private ordering, see Jennifer S Fan, “Employees as Regulators: The New Private Ordering in High Technology Companies,” Utah Law Review, no. 5 (2019): 55.
18. “可信赖算法原则和算法社会影响声明::FAT ML”，于 2018 年 10 月 1 日访问，http://www.fatml.org/resources/principles-for-accountable-algorithms。
18. “Principles for Accountable Algorithms and a Social Impact Statement for Algorithms :: FAT ML,” accessed October 1, 2018, http://www.fatml .org/resources/principles-for-accountable-algorithms.
19. 重要的批判性研究结合了历史方法，包括 Wendy Hui Kyong Chun 和 Alex Barnett 的《判别数据：相关性、邻里和新的认可政治》（马萨诸塞州剑桥：麻省理工学院出版社，2021 年）；Justin Joque 的《革命性数学：人工智能、统计学和资本主义逻辑》（纽约：Verso，2022 年）；Kate Crawford 的《人工智能地图集：权力、政治和人工智能的全球成本》（康涅狄格州纽黑文：耶鲁大学出版社，2021 年）；Meredith Broussard 的《人工智能的非智能：计算机如何误解世界》（马萨诸塞州剑桥：麻省理工学院出版社，2018 年）。有关“大数据”，请参阅先驱 Rob Kitchin 的《数据革命：大数据、开放数据、数据基础设施及其后果》（洛杉矶：SAGE 出版社，2014 年）。
19. Important critical studies incorporating historical approaches include Wendy Hui Kyong Chun and Alex Barnett, Discriminating Data: Correlation, Neighborhoods, and the New Politics of Recognition (Cambridge, MA: MIT Press, 2021); Justin Joque, Revolutionary Mathematics: Artificial Intelligence, Statistics and the Logic of Capitalism (New York: Verso, 2022); Kate Crawford, Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence (New Haven, CT: Yale University Press, 2021); Meredith Broussard, Artificial Unintelligence: How Computers Misunderstand the World (Cambridge, MA: MIT Press, 2018). For “big data,” see the pioneering Rob Kitchin, The Data Revolution: Big Data, Open Data, Data Infrastructures & Their Consequences (Los Angeles: SAGE Publications, 2014).
20. Melvin Kranzberg，《技术与历史：'Kranzberg 定律'》《技术与文化》第 27 卷第 3 期（1986 年）：第 547–48 页。
20. Melvin Kranzberg, “Technology and History: ‘Kranzberg’s Laws,’ ” Technology and Culture 27, no. 3 (1986): 547–48.
21. Enrico Coiera，“人工智能时代的医学命运”，《柳叶刀》 392，第 10162 期（2018 年 12 月 1 日）：2331，https://doi.org/10.1016/S0140-6736(18)31925–1。
21. Enrico Coiera, “The Fate of Medicine in the Time of AI,” The Lancet 392, no. 10162 (December 1, 2018): 2331, https://doi.org/10.1016/S0140 -6736(18)31925–1.
22. 有关历史作为教授伦理学的有力工具，请参阅 RR Kline，“使用历史和社会学教授工程伦理学”，IEEE 技术与社会杂志20，第 4 期（2001 年）：13-20，https://doi.org/10.1109/44.974503。
22. For history as a powerful tool in teaching ethics, see R. R. Kline, “Using History and Sociology to Teach Engineering Ethics,” IEEE Technology and Society Magazine 20, no. 4 (2001): 13–20, https://doi.org/10.1109/44 .974503.
23. 有关美国数据积累和分析不同时期的出色调查，请参阅 Dan Bouk 的《过去两个世纪个人数据的三个历史和政治经济学》，《Osiris》第 32 卷（2017 年）：第 85-106 页；Martha Hodes 的《1890 年美国人口普查中的分数和虚构》，《帝国的困扰：北美历史上的亲密关系地理》 ，安·劳拉·斯托勒主编（北卡罗来纳州达勒姆：杜克大学出版社，2006 年），第 240-70 页；Simone Browne 的《黑暗物质：关于对黑人的监视》（北卡罗来纳州达勒姆：杜克大学出版社，2015 年）；Khalil Gibran Muhammad 的《对黑人的谴责：种族、犯罪和现代城市美国的形成》（马萨诸塞州剑桥：哈佛大学出版社，2010 年）。
23. For an excellent survey of different moments of data accumulation and analysis in the US, see Dan Bouk, “The History and Political Economy of Personal Data over the Last Two Centuries in Three Acts,” Osiris 32 (2017): 85–106; Martha Hodes, “Fractions and Fictions in the United States Census of 1890,” in Haunted by Empire: Geographies of Intimacy in North American History, ed. Ann Laura Stoler (Durham, NC: Duke University Press, 2006), 240–70; Simone Browne, Dark Matters: On the Surveillance of Blackness (Durham, NC: Duke University Press, 2015); Khalil Gibran Muhammad, The Condemnation of Blackness: Race, Crime, and the Making of Modern Urban America (Cambridge, MA: Harvard University Press, 2010).
24. 莎拉·E·伊戈，《平均美国人：调查、公民和大众公众的形成》（马萨诸塞州剑桥：哈佛大学出版社，2007 年）；埃马纽埃尔·迪迪埃，《数字下的美国：量化、民主和国家统计的诞生》（马萨诸塞州剑桥：麻省理工学院出版社，2020 年）；丹尼尔·B·布克，《我们的日子如何变得有限：风险与统计个体的崛起》（伦敦：芝加哥大学出版社，2015 年）；艾米丽·克兰彻·麦钱特，《人口爆炸的形成》（纽约：牛津大学出版社，2021 年）。更广泛地说，请参阅经典研究 Geoffrey C. Bowker 和 Susan Leigh Star 的《Sorting Things Out》（马萨诸塞州剑桥：麻省理工学院出版社，1999 年）以及 Wendy Nelson Espeland 和 Michael Sauder 的《排名与反应性：公共措施如何重塑社会世界》，《美国社会学杂志》 113，第 1 期（2007 年 7 月 1 日）：1-40，https://doi.org/10.1086/517897。
24. Sarah E. Igo, The Averaged American: Surveys, Citizens, and the Making of a Mass Public (Cambridge, MA: Harvard University Press, 2007); Emmanuel Didier, America by the Numbers: Quantification, Democracy, and the Birth of National Statistics (Cambridge, MA: MIT Press, 2020); Daniel B. Bouk, How Our Days Became Numbered: Risk and the Rise of the Statistical Individual (London: University of Chicago Press, 2015); Emily Klancher Merchant, Building the Population Bomb (New York: Oxford University Press, 2021). More generally, see the classic studies Geoffrey C. Bowker and Susan Leigh Star, Sorting Things Out (Cambridge, MA: MIT Press, 1999) and Wendy Nelson Espeland and Michael Sauder, “Rankings and Reactivity: How Public Measures Recreate Social Worlds,” American Journal of Sociology 113, no. 1 (July 1, 2007): 1–40, https://doi.org/10.1086/517897.
25. Caitlin Rosenthal，《奴隶制的会计：大师与管理层》（马萨诸塞州剑桥：哈佛大学出版社，2018 年）。
25. Caitlin Rosenthal, Accounting for Slavery: Masters and Management (Cambridge, MA: Harvard University Press, 2018).
26. Theodore M. Porter，《信任数字：追求科学与公共生活中的客观性》（新泽西州普林斯顿：普林斯顿大学出版社，1995 年）。
26. Theodore M. Porter, Trust in Numbers: The Pursuit of Objectivity in Science and Public Life (Princeton, NJ: Princeton University Press, 1995).
27.参见弗兰克·帕斯夸莱（Frank Pasquale ）著《黑箱社会：控制金钱和信息的秘密算法》（马萨诸塞州剑桥：哈佛大学出版社，2015 年） 中的评论。
27. See the critique in Frank Pasquale, The Black Box Society: The Secret Algorithms That Control Money and Information (Cambridge, MA: Harvard University Press, 2015).
28. Martha Poon 强调，这些做法显然是不可避免的，必须加以解释，而不是想当然：“评分系统如何建立，如何连接、协调和互动，最重要的是，如何发展，这些细节应该与它们如何通过风险计算重新格式化和重组消费信贷行业有关。”Martha Poon，“记分卡作为消费信贷设备：Fair, Isaac & Company Incorporated 案例”，《社会学评论》 55，第 2 号补充（2007 年 10 月）：288，https://doi.org/10.1111/j.1467–954X.2007.00740.x。
28. Martha Poon stresses that the apparent inevitability of these practices is what must be explained, not taken for given: “the details of how scoring systems are made, how they connect, co-ordinate, and interact, and most of all, how they evolve, should matter in how they have reformatted and reassembled the consumer credit industry through risk calculation.” Martha Poon, “Scorecards as Devices for Consumer Credit: The Case of Fair, Isaac & Company Incorporated,” The Sociological Review 55, no. 2_suppl (October 2007): 288, https://doi.org/10.1111/j.1467–954X.2007.00740.x.
第 2 章：社会物理学和L'Homme Moyen
CHAPTER 2: SOCIAL PHYSICS AND L’HOMME MOYEN
1.摘自卡尔·皮尔逊所著《弗朗西斯·高尔顿的生平、书信和劳动》（英国剑桥：大学出版社，1914 年），第 2 卷，第 418 页，http://archive.org/details/b29000695_0002。
1. Quoted in Karl Pearson, The Life, Letters and Labours of Francis Galton (Cambridge, UK: University Press, 1914), vol. 2, 418, http://archive.org/details/b29000695_0002.
2.南丁格尔写给威廉·法尔的信，1874 年 2 月 23 日，摘自《弗洛伦斯·南丁格尔论社会与政治、哲学、科学、教育和文学：弗洛伦斯·南丁格尔文集》，第 5 卷，林恩·麦克唐纳编（安大略省滑铁卢：威尔弗里德·劳里埃大学出版社，2003 年），第 39 页。
2. Nightingale to William Farr, 23.2.1874, in Florence Nightingale on Society and Politics, Philosophy, Science, Education and Literature: Collected Works of Florence Nightingale, Volume 5, ed. Lynn McDonald (Waterloo, ON: Wilfrid Laurier University Press, 2003), 39.
3. Ian Hacking，《驯服机遇》（英国剑桥：剑桥大学出版社，1990 年），第 106 页。
3. Ian Hacking, The Taming of Chance (Cambridge, UK: Cambridge University Press, 1990), 106.
4. Joseph Lottin，Quetelet，Statisticien et Sociologue（鲁汶：高等哲学研究所，1912），52； Theodore M. Porter，统计思维的兴起，1820-1900（新泽西州普林斯顿：普林斯顿大学出版社，1986），47。
4. Joseph Lottin, Quetelet, Statisticien et Sociologue (Louvain: Institut supérieur de philosophie, 1912), 52; Theodore M. Porter, The Rise of Statistical Thinking, 1820–1900 (Princeton, NJ: Princeton University Press, 1986), 47.
5. David Aubin，“易于应用于社会的力学原理：阿道夫·凯特勒未发表的笔记，揭示了他的社会物理学根源”，《数学史》第 41 卷第 2 期（2014 年 5 月 1 日）：209、216，https://doi.org/10.1016/j.hm.2014.01.001。
5. David Aubin, “Principles of Mechanics That Are Susceptible of Application to Society: An Unpublished Notebook of Adolphe Quetelet at the Root of His Social Physics,” Historia Mathematica 41, no. 2 (May 1, 2014): 209, 216, https://doi.org/10.1016/j.hm.2014.01.001.
6. Kevin Donnelly，Adolphe Quetelet，《社会物理学与普通科学家，1796–1874》（Routledge，2015），73，https://doi.org/10.4324/9781315653662。
6. Kevin Donnelly, Adolphe Quetelet, Social Physics and the Average Men of Science, 1796–1874 (Routledge, 2015), 73, https://doi.org/10 .4324/9781315653662.
7.译自保罗·F·拉扎斯菲尔德著《社会学量化史笔记——趋势、来源和问题》，《Isis》 52卷2期（1961年）：第293页。
7. Translated in Paul F. Lazarsfeld, “Notes on the History of Quantification in Sociology—Trends, Sources and Problems,” Isis 52, no. 2 (1961): 293.
8.参见 Morgane Labbé，“L'arithmétique politique en Allemagne au début du 19e siècle：receptions et polémiques”，Journal Electronique d'Histoire des Probabilités et de la Statistique 4，第 4 期。 1（2008）：7。
8. See Morgane Labbé, “L’arithmétique politique en Allemagne au début du 19e siècle: réceptions et polémiques,” Journal Electronique d’Histoire des Probabilités et de la Statistique 4, no. 1 (2008): 7.
9.佩吉·努南，《他们已经失去了那种爱的感觉》，《华尔街日报》，2012 年 11 月 20 日访问，http://online.wsj.com/article/SB10001424053111904800304576474620336602248.html。
9. Peggy Noonan, “They’ve Lost That Lovin’ Feeling,” Wall Street Journal, accessed November 20, 2012, http://online.wsj.com/article/SB10001424053 111904800304576474620336602248.html.
10. 有关 17 世纪以来发展情况的有力考察，请参阅大量文献中的 Jacqueline Wernimont 的《数字生命：量子媒体中的生与死》（马萨诸塞州剑桥：麻省理工学院出版社，2018 年），尤其是第 2 章；Andrea Rusnock 的《量化、精确度和准确度：旧制度下的人口决定》，《精确度的价值》，M. Norton Wise 主编（新泽西州普林斯顿：普林斯顿大学出版社，1995 年），第 17-38 页；William Deringer 的《计算价值：金融、政治和量化时代》（马萨诸塞州剑桥：哈佛大学出版社，2018 年），他强调了更早的数字思维的非政府起源。
10. For powerful examinations of developments in the seventeenth century forward, see, among an extensive literature, Jacqueline Wernimont, Numbered Lives: Life and Death in Quantum Media (Cambridge, MA: MIT Press, 2018), esp. ch. 2; Andrea Rusnock, “Quantification, Precision, and Accuracy: Determinations of Population in the Ancien Régime,” in Values of Precision, ed. M. Norton Wise (Princeton, NJ: Princeton University Press, 1995), 17–38; William Deringer, Calculated Values: Finance, Politics, and the Quantitative Age (Cambridge, MA: Harvard University Press, 2018), who stresses the nongovernmental origins of much earlier numerical thinking.
11. Lisa Gitelman 编，《原始数据是一种矛盾》（马萨诸塞州剑桥：麻省理工学院出版社，2013 年）。几本主要的论文集在不同时间和全球范围内发展了这些思想：Elena Aronova、Christina von Oertzen 和 David Sepkoski 编，《数据历史》，《Osiris》第 32 卷（芝加哥：芝加哥大学出版社，2017 年）；Soraya de Chadarevian 和 Theodore M. Porter，《数据和数据库的历史》，第 48 卷，第 5 期，《自然科学历史研究》 ，2018 年；另请参阅 Amelia Acker，“走向数据解释学”，《计算史年鉴》，IEEE第 37 卷，第 3 期（2015 年）：第 70-75 页。
11. Lisa Gitelman, ed., Raw Data Is an Oxymoron. (Cambridge, MA: MIT Press, 2013). Several major collections of essays develop these ideas across time and the globe: Elena Aronova, Christina von Oertzen, and David Sepkoski, eds., Data Histories, Osiris 32 (Chicago: University of Chicago Press, 2017); Soraya de Chadarevian and Theodore M. Porter, Histories of Data and the Database, vol. 48, no. 5, Historical Studies in the Natural Sciences, 2018; see also Amelia Acker, “Toward a Hermeneutics of Data,” Annals of the History of Computing, IEEE 37, no. 3 (2015): 70–75.
12. 黑客，《机会的驯服》，2。
12. Hacking, The Taming of Chance, 2.
13. 有关中国细致入微的描述，请参阅 Tong Lam 的《对事实的热情：社会调查与中国民族国家的建设，1900-1949 年》（伯克利：加州大学出版社，2011 年），第 1 章。
13. For a nuanced account of China, see Tong Lam, A Passion for Facts: Social Surveys and the Construction of the Chinese Nation State, 1900–1949. (Berkeley: University of California Press, 2011), ch. 1.
14. Jacqueline Wernimont，《Numbered Lives》，第 28 页；有关收集此数据的劳动，请参阅 Deborah E. Harkness 的《街头观点：伊丽莎白时代伦敦的妇女和医疗工作》，《医学史公报》第 82 卷，第 1 期（2008 年）：第 52-85 页。
14. Jacqueline Wernimont, Numbered Lives, 28; for the labor of collecting this data, see Deborah E. Harkness, “A View from the Streets: Women and Medical Work in Elizabethan London,” Bulletin of the History of Medicine 82, no. 1 (2008): 52–85.
15. Victor L. Hilts，“Aliis Exterendum，或伦敦统计协会的起源”，Isis 69，第 15 期。 1（1978 年 3 月）：21-43，https://doi.org/10.1086/351931。
15. Victor L. Hilts, “Aliis Exterendum, or, the Origins of the Statistical Society of London,” Isis 69, no. 1 (March 1978): 21–43, https://doi.org/10.1086/351931.
16. 论证是威廉·德林格 (William Deringer) 的《计算价值》 (Calculated Values) 的核心焦点。
16. The focus on argument is central in William Deringer, Calculated Values.
17. 参见 David Sepkoski 和 Marco Tamborini，“‘科学的形象’：十九世纪的光武主义、统计学和自然史的视觉语言”，《自然科学历史研究》第 48 卷第 1 期（2018 年 2 月 1 日）：56–109，https://doi.org/10.1525/hsns.2018.48.1.56；Deringer，《计算值》。
17. See David Sepkoski and Marco Tamborini, “ ‘An Image of Science’: Cameralism, Statistics, and the Visual Language of Natural History in the Nineteenth Century,” Historical Studies in the Natural Sciences 48, no. 1 (February 1, 2018): 56–109, https://doi.org/10.1525/hsns.2018.48.1.56; Deringer, Calculated Values.
18. 参见 Jean-Guy Prévost 和 Jean-Pierre Beaud，《统计、公共辩论和国家，1800–1945：数字的社会、政治和思想史》（劳特利奇，2016 年），第 3 页。
18. See Jean-Guy Prévost and Jean-Pierre Beaud, Statistics, Public Debate and the State, 1800–1945: A Social, Political and Intellectual History of Numbers (Routledge, 2016), 3.
19. 凯特勒的国际工作是唐纳利（Donnelly），阿道夫·凯特勒（Adolphe Quetelet），《社会物理学和普通科学家，1796–1874》的一个主要主题。
19. Quetelet’s international work is a major theme of Donnelly, Adolphe Quetelet, Social Physics and the Average Men of Science, 1796–1874.
20. Adolphe Quetelet，《人及其能力发展论》（爱丁堡：W. and R. Chambers，1842 年），第 6 页，http://archive.org/details/treatise onmandev00quet。
20. Adolphe Quetelet, A Treatise on Man and the Development of His Faculties (Edinburgh: W. and R. Chambers, 1842), 6, http://archive.org/details/treatise onmandev00quet.
21. 凯特勒，6。
21. Quetelet, 6.
22. David Aubin，“论数学作为天文台工具和仪器的认识论和社会基础，1793-1846”，载Johannes Lenhard 和 Martin Carrier 编《数学作为一种工具》，第 327 卷（瑞士 Cham：Springer International Publishing，2017 年），第 290-91 页，https://doi.org/10.1007/978-3-319-54469-4_10；Porter， 《统计思维的兴起，1820-1900》，第 42 页；Donnelly，Adolphe Quetelet，《社会物理学和普通科学家，1796-1874》，第 111-12 页。Aubin 强调观察数据的技术；Donnelly 强调 Quetelet 越来越有能力从欧洲各地的来源获取数据。
22. David Aubin, “On the Epistemic and Social Foundations of Mathematics as Tool and Instrument in Observatories, 1793–1846,” in Mathematics as a Tool, ed. Johannes Lenhard and Martin Carrier, vol. 327 (Cham, Switzerland: Springer International Publishing, 2017), 290–91, https://doi.org/10 .1007/978–3–319–54469–4_10; Porter, The Rise of Statistical Thinking, 18201900, 42; Donnelly, Adolphe Quetelet, Social Physics and the Average Men of Science, 1796–1874, 111–12. Aubin stresses the techniques of observing data; Donnelly emphasizes Quetelet’s increasing ability to get data from sources across Europe.
23. Hacking，《机会的驯服》，109。
23. Hacking, The Taming of Chance, 109.
24. Adolphe Quetelet，Recherches Statistiques（布鲁塞尔：M. Hayez，1844），54。
24. Adolphe Quetelet, Recherches Statistiques (Brussels: M. Hayez, 1844), 54.
25. Hacking，《机会的驯服》，107。
25. Hacking, The Taming of Chance, 107.
26. 凯特勒，《人论》，第5页。
26. Quetelet, A Treatise on Man, 5.
27. 凯特勒，6。
27. Quetelet, 6.
28. 凯特勒，6。
28. Quetelet, 6.
29. 凯特勒，6。
29. Quetelet, 6.
30. 凯特勒，6。
30. Quetelet, 6.
31. Hacking，《机会的驯服》，108。
31. Hacking, The Taming of Chance, 108.
32. 玛格丽特·撒切尔，《女性自己》采访，1987 年 9 月 23 日，https://www.margaretthatcher.org/document/106689。
32. Margaret Thatcher, Interview for Woman’s Own, 23.9.1987, https://www .margaretthatcher.org/document/106689.
33. Porter，《统计思维的兴起，1820-1900》，55。
33. Porter, The Rise of Statistical Thinking, 1820–1900, 55.
34. 凯特勒，《人论》，第7页。
34. Quetelet, A Treatise on Man, 7.
35. Porter，《统计思维的兴起，1820-1900》，第46页。
35. Porter, The Rise of Statistical Thinking, 1820–1900, 46.
36. 波特，104。
36. Porter, 104.
37. 黑客，驯服机会，108。
37. Hacking, Taming of Chance, 108.
38. Adrian Wooldridge，《测量心智：1860-1990年左右的英国教育与心理学》（纽约：剑桥大学出版社，1994年），第74页。
38. Adrian Wooldridge, Measuring the Mind: Education and Psychology in England, c. 1860–c. 1990 (New York: Cambridge University Press, 1994), 74.
39. 引自皮尔逊著《弗朗西斯·高尔顿的生平、书信和劳动》，第 2 卷，第 419 页。
39. Quoted in Pearson, The Life, Letters and Labours of Francis Galton, v. 2, 419.
40. 引自 Pearson，第 2 卷，419。
40. Quoted in Pearson, v. 2, 419.
第三章：异常者的统计数据
CHAPTER 3: THE STATISTICS OF THE DEVIANT
1.弗洛伦斯·南丁格尔，《影响英国军队健康、效率和医院管理事项笔记》（伦敦：哈里森父子公司，1858 年），第 518 页。
1. Florence Nightingale, Notes on Matters Affecting the Health, Efficiency and Hospital Administration of the British Army (London: Harrison and Sons, 1858), 518.
2.弗朗西斯·高尔顿，《遗传、天赋和性格》，《麦克米伦杂志》 12（1865）：166。
2. Francis Galton, “Heredity Talent And Character,” Macmillan’s Magazine 12 (1865): 166.
3.高尔顿解释说，达尔文“在我的思想发展中开创了一个显著的时代，就像在人类思想的发展中一样。”弗朗西斯·高尔顿爵士，《我一生的回忆》（纽约：达顿，1909 年），第 287 页。
3. Galton explained that Darwin “made a marked epoch in my own mental development, as it did in that of human thought generally.” Sir Francis Galton, Memories of My Life (New York: Dutton, 1909), 287.
4. Galton，《遗传、天赋和性格》，第157页。
4. Galton, “Heredity Talent And Character,” 157.
5.高尔顿，165。
5. Galton, 165.
6. Chris Renwick，“从政治经济学到社会学：弗朗西斯·高尔顿与优生学的社会科学起源”，《英国科学史杂志》第 44 卷第 3 期（2011 年 9 月）：352，https://doi.org/10.1017/S000 7087410001524。
6. Chris Renwick, “From Political Economy to Sociology: Francis Galton and the Social-Scientific Origins of Eugenics,” The British Journal for the History of Science 44, no. 3 (September 2011): 352, https://doi.org/10.1017/S000 7087410001524.
7.弗朗西斯·高尔顿，《遗传天才：对其规律和后果的探究》（伦敦：麦克米伦，1869 年），第 14 页，http://archive.org/details/hereditary genius1869galt。
7. Francis Galton, Hereditary Genius: An Inquiry into Its Laws and Consequences (London: Macmillan, 1869), 14, http://archive.org/details/hereditary genius1869galt.
8.罗斯，引自托马斯·伦纳德著《不自由的改革者：进步时代的种族、优生学和美国经济学》（新泽西州普林斯顿：普林斯顿大学出版社，2016 年），第 110 页。
8. Ross, quoted in Thomas C. Leonard, Illiberal Reformers: Race, Eugenics, and American Economics in the Progressive Era (Princeton, NJ: Princeton University Press, 2016), 110.
9.有关优生学与当代数据实践之间的联系，请参阅 Chun 和 Barnett 的《Discriminating Data》，第 1 章。
9. For connections between this eugenics and contemporary data practices, see Chun and Barnett, Discriminating Data, ch. 1.
10. Alain Desrosières，《大数政治：统计推理史》（马萨诸塞州剑桥：哈佛大学出版社，1998 年），第 113 页；Stephen M. Stigler，《统计史：1900 年前的不确定性测量》（马萨诸塞州剑桥：哈佛大学出版社贝尔纳普出版社，1986 年），第 271 页。
10. Alain Desrosières, The Politics of Large Numbers: A History of Statistical Reasoning (Cambridge, MA: Harvard University Press, 1998), 113; Stephen M. Stigler, The History of Statistics: The Measurement of Uncertainty Before 1900 (Cambridge, MA: The Belknap Press of Harvard University Press, 1986), 271.
11. 弗朗西斯·高尔顿，《典型的遗传规律》，英国皇家学会。《会员会议记录通知》 8（1877 年 2 月 16 日）：第 291 页。
11. Francis Galton, “Typical Laws of Heredity,” Royal Institution of Great Britain. Notices of the Proceedings at the Meetings of the Members 8 (February 16, 1877): 291.
12. 弗朗西斯·高尔顿 (Francis Galton)，《人体测量实验室；由弗朗西斯·高尔顿 (Francis Galton)，FRS 编纂，用于测定身高、体重、跨度、呼吸能力、拉挤强度、吹气速度、听觉、视觉、色觉和其他个人数据》 (伦敦：William Clowes，1884 年)，3，http://archive.org/details/b30579132。
12. Francis Galton, Anthropometric Laboratory; Arranged by Francis Galton, FRS, for the Determination of Height, Weight, Span, Breathing Power, Strength of Pull and Squeeze, Quickness of Blow, Hearing, Seeing, Colour-Sense, and Other Personal Data (London: William Clowes, 1884), 3, http:// archive. org/details/b3 0579132.
13. 高尔顿，4。
13. Galton, 4.
14. 库尔特·丹齐格，《建构主体：心理学研究的历史起源》（英国剑桥：剑桥大学出版社，1990年），57。
14. Kurt Danziger, Constructing the Subject: Historical Origins of Psychological Research (Cambridge, UK: Cambridge University Press, 1990), 57.
15. 丹齐格，77。
15. Danziger, 77.
16. 丹齐格，110。
16. Danziger, 110.
17. Porter，《统计思维的兴起，1820-1900》，第311页。
17. Porter, The Rise of Statistical Thinking, 1820–1900, 311.
18. Porter，第 304–5 页。“从另一个意义上说，皮尔逊是凯特勒的忠实追随者。两人都认同数字的普遍性和不连续性。两人都认为科学的任务不是绘制大胆的新当然，而要研究社会发展的规律，以便科学的政策能够确认这些规律，并消除实现这些规律的一切障碍。”
18. Porter, 304–5. “In another sense, Pearson was a true follower of Quetelet. Both agreed on the universality of numbers and on the absence of discontinuity. Both maintained that the task of science was not to chart a bold new course, but to study the laws of social development so that scientific policy might affirm them and remove all obstacles to their attainment.”
19. Theodore M. Porter，《卡尔·皮尔逊：统计时代的科学生活》（新泽西州普林斯顿：普林斯顿大学出版社，2004 年），第 261 页。
19. Theodore M. Porter, Karl Pearson: The Scientific Life in a Statistical Age (Princeton, NJ: Princeton University Press, 2004), 261.
20. Karl Pearson，《论人类的遗传规律：II。论人类心智和道德特征的遗传及其与身体特征遗传的比较》，《Biometrika》 3，第 2/3 期（1904 年）：136，https://doi.org/10.2307/2331479。
20. Karl Pearson, “On the Laws of Inheritance in Man: II. On the Inheritance of the Mental and Moral Characters in Man, and Its Comparison with the Inheritance of the Physical Characters,” Biometrika 3, no. 2/3 (1904): 136, https://doi.org/10.2307/2331479.
21. M. Eileen Magnello，“生物识别与优生学的不相关性：Karl Pearson 在伦敦大学学院职业生涯中的实验室工作竞争形式，第一部分”，《科学史》第 37 卷，第 1 期（1999 年 3 月 1 日）：79–106，https://doi.org/10.1177/007327539903700103；M. Eileen Magnello，“生物识别与优生学的不相关性：Karl Pearson 在伦敦大学学院职业生涯中的实验室工作竞争形式，第二部分”，《科学史》第 37 卷，第 2 期（1999 年 6 月 1 日）：123–50。
21. M. Eileen Magnello, “The Non-Correlation of Biometrics and Eugenics: Rival Forms of Laboratory Work in Karl Pearson’s Career at University College London, Part 1,” History of Science 37, no. 1 (March 1, 1999): 79–106, https://doi.org/10.1177/007327539903700103; M. Eileen Magnello, “The Non-Correlation of Biometrics and Eugenics: Rival Forms of Laboratory Work in Karl Pearson’s Career at University College London, Part 2,” History of Science 37, no. 2 (June 1, 1999): 123–50.
22.Porter ，Karl Pearson：《统计时代的科学生活》，263。
22. Porter, Karl Pearson: The Scientific Life in a Statistical Age, 263.
23. Pearson，“论人类继承法”，第 136 页。有关女性作为计算机，请参阅 Jennifer S. Light，“当计算机是女性时”，《科技与文化》第 40 卷第 3 期（1999 年）：第 455-483 页。
23. Pearson, “On the Laws of Inheritance in Man,” 136. On women as computers, see Jennifer S. Light, “When Computers Were Women,” Technology and Culture 40, no. 3 (1999): 455–83.
24. David Alan Grier， 《当计算机是人类时》（新泽西州普林斯顿：普林斯顿大学出版社，2005 年），111。
24. David Alan Grier, When Computers Were Human (Princeton, NJ: Princeton University Press, 2005), 111.
25. 转引自Grier，第117页。
25. Quoted in Grier, 117.
26. Pearson，《弗朗西斯·高尔顿的生平、书信和劳动》，IIIA：305。
26. Pearson, The Life, Letters and Labours of Francis Galton, IIIA: 305.
27. Pearson，《论人类继承规律》，第159页。
27. Pearson, “On the Laws of Inheritance in Man,” 159.
28. Alice Lee 和 Karl Pearson，《人类进化问题的数据。VI。人类头骨相关性的初步研究》，《伦敦皇家学会哲学学报》。A 系列，包含数学或物理论文196（1901 年）：259。
28. Alice Lee and Karl Pearson, “Data for the Problem of Evolution in Man. VI. A First Study of the Correlation of the Human Skull,” Philosophical Transactions of the Royal Society of London. Series A, Containing Papers of a Mathematical or Physical Character 196 (1901): 259.
29. Karl Pearson，《论人类心智与道德品格的遗传及其与身体品格遗传的比较》，《英国及爱尔兰人类学研究所期刊》第 33 卷（1903 年）：第 207 页，https://doi.org/10.2307/2842809。
29. Karl Pearson, “On the Inheritance of the Mental and Moral Characters in Man, and Its Comparison with the Inheritance of the Physical Characters,” The Journal of the Anthropological Institute of Great Britain and Ireland 33 (1903): 207, https://doi.org/10.2307/2842809.
30. Karl Pearson 和 Margaret Moul，“通过对俄罗斯和波兰犹太儿童的调查说明外来移民进入英国的问题”，优生学年鉴1，第 1 期（1925 年）：7，https://doi.org/10.1111/j.1469–1809.1925.tb02037.x。
30. Karl Pearson and Margaret Moul, “The Problem of Alien Immigration into Great Britain, Illustrated by an Examination of Russian and Polish Jewish Children,” Annals of Eugenics 1, no. 1 (1925): 7, https://doi.org/10.1111/j .1469–1809.1925.tb02037.x.
31. Karl Pearson，《死亡机会及其他进化研究》（伦敦、纽约：E. Arnold，1897 年），104，http://archive.org/details/cu31924097311579。转引自 Porter 的《Karl Pearson：统计时代的科学生活》，267。
31. Karl Pearson, The Chances of Death, and Other Studies in Evolution (London, New York: E. Arnold, 1897), 104, http://archive.org/details/cu3192 4097311579. Quoted in Porter, Karl Pearson: The Scientific Life in a Statistical Age, 267.
32. Karl Pearson 和 Ethel M. Elderton，《父母酗酒对子女体格和能力影响的第二次研究：对第一部回忆录中某些医学批评者的回应以及对他们所引用的反驳证据的审查》（伦敦，Dulau and Co.，1910 年），第 34 页，http://archive.org/details/secondstudyofinf00pear；讨论见 PC Mahalanobis 的《Karl Pearson，1857–1936》，《Sankhyā：印度统计学杂志》 2，第 4 期（1936 年）：第 368 页；另请参阅 Donald A. MacKenzie 的《英国的统计学，1865–1930：科学知识的社会建构》（爱丁堡：爱丁堡大学出版社，1981 年），第 139 页。
32. Karl Pearson and Ethel M. Elderton, A Second Study of the Influence of Parental Alcoholism on the Physique and Ability of the Offspring: Being a Reply to Certain Medical Critics of the First Memoir and an Examination of the Rebutting Evidence Cited by Them (London, Dulau and Co., 1910), 34, http://archive.org/details/secondstudyofinf00pear; discussed in P. C. Mahalanobis, “Karl Pearson, 1857–1936,” Sankhyā: The Indian Journal of Statistics 2, no. 4 (1936): 368; see also Donald A. MacKenzie, Statistics in Britain, 1865–1930: The Social Construction of Scientific Knowledge (Edinburgh: Edinburgh University Press, 1981), 139.
33. Michel Armatte，“发明与干预统计。Une卡尔·皮尔逊会议范例 (1912)，” Politix。Revue des sciences Sociales du politique 7，第 25 期 (1994)：30，https://doi.org/10.3406/polix.1994.1823。
33. Michel Armatte, “Invention et intervention statistiques. Une conférence exemplaire de Karl Pearson (1912),” Politix. Revue des sciences sociales du politique 7, no. 25 (1994): 30, https://doi.org/10.3406/polix.1994.1823.
34. Porter，《统计思维的兴起，1820-1900》，第298页。
34. Porter, The Rise of Statistical Thinking, 1820–1900, 298.
35. Pearson，《弗朗西斯·高尔顿的生平、书信和劳动》，IIIa：57。
35. Pearson, The Life, Letters and Labours of Francis Galton, IIIa:57.
36. Robert A. Nye，“优生学帝国的崛起与衰落：关于生物医学思想对现代社会影响的最新观点”，《历史期刊》第 36 卷，第 3 期（1993 年 9 月）：695，https://doi.org/10.1017/S00 18246X00014369。
36. Robert A. Nye, “The Rise and Fall of the Eugenics Empire: Recent Perspectives on the Impact of Biomedical Thought in Modern Society,” The Historical Journal 36, no. 3 (September 1993): 695, https://doi.org/10.1017/S00 18246X00014369.
37. Brajendranath Seal，《种族、部落、民族的意义》，《种族间问题论文集》，提交给 1911 年 7 月 26-29 日在伦敦大学举行的第一届世界种族大会》， Gustav Spiller 主编（伦敦：PS King & Son；波士顿，世界和平基金会，1911 年），第 1 页，http://archive.org/details/papersoninterrac00univiala；我们的评论非常感谢 Projit Bihari Mukharji，《孟加拉法老：高种姓雅利安主义、泛埃及主义和 20 世纪孟加拉生物特征民族主义的争议历史》，《社会与历史比较研究》第 59 卷，第 1 期。 2 (2017 年 4 月)：450，https://doi.org/10.1017/S001041751700010X。有关其更广泛的统计计划，请参阅 Theodora Dryer，“设计确定性：1920-1970 年焦虑时代算法计算的兴起”(博士论文，加州大学圣地亚哥分校，20194)，157-162；并参阅即将出版的作品 Sananda Sahoo，“马哈拉诺比斯的生物特征数据的多重生命作为生物价值传到印度福利国家”(正在审查中)。
37. Brajendranath Seal, “Meaning of Race, Tribe, Nation,” in Papers on Inter-Racial Problems, Communicated to the First Universal Races Congress, Held at the University of London, July 26–29, 1911, ed. Gustav Spiller (London: P. S. King & Son; Boston, The World’s Peace Foundation, 1911), 1, http://archive.org/details/papersoninterrac00univiala; our remarks much indebted to Projit Bihari Mukharji, “The Bengali Pharaoh: Upper-Caste Aryanism, Pan-Egyptianism, and the Contested History of Biometric Nationalism in Twentieth-Century Bengal,” Comparative Studies in Society and History 59, no. 2 (April 2017): 450, https://doi.org/10.1017/ S001041751700010X. For his broader statistical program, see Theodora Dryer, “Designing Certainty: The Rise of Algorithmic Computing in an Age of Anxiety 1920–1970” (PhD Thesis, UC San Diego, 20194), 157–162; and see the forthcoming work Sananda Sahoo, “Multiple lives of Mahalanobis’ biometric data travel as biovalue to India’s welfare state” (under review).
38. Seal，《种族、部落、民族的意义》，第2页。
38. Seal, “Meaning of Race, Tribe, Nation,” 2.
39. 印章，3。
39. Seal, 3.
40. Nikhil Menon，“‘花哨的计算机’：独立印度的计算机和规划”，《现代亚洲研究》第 52 卷，第 2 期（2018 年 3 月）：第 421–57 页，https://doi.org/10.1017/S0026749X16000135；Sandeep Mertia，“马哈拉诺比斯梦见过机器人吗？”，《数据生活：印度计算文化论文集》，Sandeep Mertia 和 Ravi Sundaram 主编（阿姆斯特丹：网络文化研究所，2020 年），第 26–33 页。
40. Nikhil Menon, “ ‘Fancy Calculating Machine’: Computers and Planning in Independent India,” Modern Asian Studies 52, no. 2 (March 2018): 421–57, https://doi.org/10.1017/S0026749X16000135; Sandeep Mertia, “Did Mahalanobis Dream of Androids?,” in Lives of Data: Essays on Computational Cultures from India, ed. Sandeep Mertia and Ravi Sundaram (Amsterdam: Institute of Network Cultures, 2020), 26–33.
41. 关于这一观点，请参阅 Projit Bihari Mukharji 的《剖析轮廓镜：种族技术的面部化和两次世界大战期间英属印度生物识别民族主义的兴起》，《历史与技术》第 31 卷第 4 期（2015 年 10 月 2 日）：第 392 页，https://doi.org/10.1080/07341512.2015.1127459。“剖析轮廓镜让我们得以一窥生物识别技术在南亚乃至更广泛领域发展的重要时刻。它向我们展示了两次世界大战期间，民族主义者而非衰弱的殖民国家开发了生物识别技术。”
41. For this idea, see Projit Bihari Mukharji, “Profiling the Profiloscope: Facialization of Race Technologies and the Rise of Biometric Nationalism in Inter-War British India,” History and Technology 31, no. 4 (October 2, 2015): 392, https://doi.org/10.1080/07341512.2015.1127459. “Profiling the Profiloscope allows us then to glimpse an important moment in the development of biometric technologies in South Asia in particular, but also more generally. It shows us that in the interwar period, nationalists rather than the weakened colonial state developed biometric technologies.”
42. PC Mahalanobis，《孟加拉人种混合分析》，《孟加拉亚洲学会刊》第 23 卷（1927 年）：第 323 页。
42. P. C. Mahalanobis, “Analysis of Race-Mixture in Bengal,” Journal of the Asiatic Society of Bengal 23 (1927): 323.
43. PC Mahalanobis 等人，《1941 年联合省人体测量调查：一项统计研究》，《Sankhyā：印度统计杂志》第 9 卷，第 2/3 期（1949 年）：第 168 页。
43. P. C. Mahalanobis et al., “Anthropometric Survey of the United Provinces, 1941: A Statistical Study,” Sankhyā: The Indian Journal of Statistics 9, no. 2/3 (1949): 168.
44.Mahalanobis 等人，180。
44. Mahalanobis et al., 180.
45. WE Burghardt Du Bois，“第一届全球种族会议论文主要结论摘要”，系列 1，盒子 007，马萨诸塞大学阿默斯特分校图书馆特藏和大学档案馆，https://www.digitalcommonwealth.org/search/commonwealth-oai:h128q9079。
45. W. E. Burghardt Du Bois, “A Summary of the Main Conclusions of the Papers Presented to the First Universal Races Conference,” Series 1, Box 007, Special Collections and University Archives, University of Massachusetts Amherst Libraries, https://www.digitalcommonwealth.org/search/ commonwealth-oai:h128q9079.
第四章：数据、情报和政策
CHAPTER 4: DATA, INTELLIGENCE, AND POLICY
1. Frederick L. Hoffman，《美国黑人的种族特征和倾向》（美国经济协会出版物，1896 年），第 2 页，http://archive.org/details/jstor-2560438。有关 Hoffman 的内容，请参阅 Daniel B. Bouk，《我们的日子如何变得无多：风险与统计个体的崛起》（伦敦：芝加哥大学出版社，2015 年），第 48–52 页。
1. Frederick L. Hoffman, The Race Traits and Tendencies of the American Negro (Publications of the American Economic Association, 1896), 2, http://archive.org/details/jstor-2560438. For Hoffman, see Daniel B. Bouk, How Our Days Became Numbered: Risk and the Rise of the Statistical Individual (London: University of Chicago Press, 2015), 48–52.
2.霍夫曼，《美国黑人的种族特征和倾向》，312。
2. Hoffman, The Race Traits and Tendencies of the American Negro, 312.
3.参见 Beatrix Hoffman，《科学种族主义、保险和对福利国家的反对：弗雷德里克·L·霍夫曼的跨大西洋之旅》，《镀金时代和进步时代杂志》第 2 卷，第 2 期（2003 年 4 月）：150–90 页，https://doi.org/10.1017/S1537781400002450。
3. See Beatrix Hoffman, “Scientific Racism, Insurance, and Opposition to the Welfare State: Frederick L. Hoffman’s Transatlantic Journey,” The Journal of the Gilded Age and Progressive Era 2, no. 2 (April 2003): 150–90, https:// doi.org/10.1017/S1537781400002450.
4. WE Burghardt Du Bois，《<美国黑人的种族特征和倾向>评论》，《美国政治和社会科学院年鉴》第9卷，第1期（1897年）：129页。
4. W. E. Burghardt Du Bois, “Review of ‘Race Traits and Tendencies of the American Negro,’ ” The Annals of the American Academy of Political and Social Science 9, no. 1 (1897): 129.
5. Ayah Nurddin，《优生学的黑人政治》，Nursing Clio（博客），2017 年 6 月 1 日，https://nursingclio.org/2017/06/01/the-black-politics-of-eugenics/。
5. Ayah Nurddin, “The Black Politics of Eugenics,” Nursing Clio (blog), June 1, 2017, https://nursingclio.org/2017/06/01/the-black-politics-of-eugenics/.
6.乔治·M·弗雷德里克森，《白人心灵中的黑人形象：1817-1914 年非裔美国人性格与命运之争》（宾夕法尼亚州斯克兰顿：Harper & Row 出版，1987 年），第 249 页。
6. George M. Fredrickson, The Black Image in the White Mind: The Debate on Afro-American Character and Destiny, 1817–1914 (Scranton, PA: Distributed by Harper & Row, 1987), 249.
7. Khalil Gibran Muhammad，《对黑人的谴责：种族、犯罪和现代城市美国的形成》（马萨诸塞州剑桥：哈佛大学出版社，2010 年），第 5 页。
7. Khalil Gibran Muhammad, The Condemnation of Blackness: Race, Crime, and the Making of Modern Urban America (Cambridge, MA: Harvard University Press, 2010), 5.
8.Desrosières，《大数政治》，139。
8. Desrosières, The Politics of Large Numbers, 139.
9. TS Simey，Charles Booth，社会科学家（伦敦，1960 年），48，http://hdl.handle.net/2027/uc1.b3620533。
9. T. S. Simey, Charles Booth, Social Scientist (London, 1960), 48, http://hdl .handle.net/2027/uc1.b3620533.
10. Charles Booth，《英格兰和威尔士的贫困老人》（伦敦：Macmillan and Co.，1894 年），423，http://archive.org/details/agedpoorinengla00bootgoog。
10. Charles Booth, The Aged Poor in England and Wales (London: Macmillan and Co., 1894), 423, http://archive.org/details/agedpoorinengla00bootgoog.
11. 比较斯蒂格勒，《统计史》，第354页。
11. Compare Stigler, The History of Statistics, 354.
12. 卡尔·皮尔逊，《科学语法》，第3版（伦敦：Adam & Charles Black，1911），157。
12. Karl Pearson, The Grammar of Science, 3rd ed. (London: Adam & Charles Black, 1911), 157.
13. G. Udny Yule，“论相关理论”，《皇家统计学会杂志》第 60 卷，第 4 期（1897 年 12 月）：812，https://doi.org/10.2307/2979746。
13. G. Udny Yule, “On the Theory of Correlation,” Journal of the Royal Statistical Society 60, no. 4 (December 1897): 812, https://doi.org/10.2307/2979746.
14. G. Udny Yule，“对英格兰贫困化变化原因的调查，主要是在最近两个人口普查间隔期（第一部分）”，《皇家统计学会杂志》第 62 卷，第 2 期（1899 年）：249，https://doi.org/10.2307/2979889。
14. G. Udny Yule, “An Investigation into the Causes of Changes in Pauperism in England, Chiefly During the Last Two Intercensal Decades (Part I.),” Journal of the Royal Statistical Society 62, no. 2 (1899): 249, https://doi .org/10.2307/2979889.
15. 圣诞节，250页。
15. Yule, 250.
16. G. Udny Yule，“论总贫困率与救济比例之间的关系”，《经济学期刊》第 5 卷，第 20 期（1895 年）：第 605 页，https://doi.org/10.2307/2956650；讨论见 C. Terence Mills 所著《乔治·乌德尼·尤尔的统计传记：一个浪子的世界》（纽卡斯尔：剑桥学者出版社，2017 年），第 43 页。
16. G. Udny Yule, “On the Correlation of Total Pauperism with Proportion of Out-Relief,” The Economic Journal 5, no. 20 (1895): 605, https://doi.org/10 .2307/2956650; discussed in C. Terence Mills, A Statistical Biography of George Udny Yule: A Loafer of the World (Newcastle upon Tyne: Cambridge Scholars Publisher, 2017), 43.
17. G. Udny Yule，“论贫困总水平与救济比例之间的关系”，《经济学期刊》 5，第 20 期（1895 年）：606，https://doi.org/10.2307/2956650。
17. G. Udny Yule, “On the Correlation of Total Pauperism with Proportion of Out-Relief,” The Economic Journal 5, no. 20 (1895): 606, https://doi.org/10 .2307/2956650.
18. 他进一步指出，“详细的知识”可能提供一些因果理解：“详细的知识有时可能使人说‘这里的贫困率很低，因为外救济的比例很小’，或者‘外救济的比例很大，因为’高度贫困和联盟的其他工业条件’：但这种情况将是例外，并且通常仅指与平均值的较大偏差。” Yule，《论总贫困与外部救济比例的相关性》，1895 年，605n2；在 Mills，《统计传记》，46 中进行了讨论。
18. He went further to note that “detailed knowledge” might give some causal understanding: “Detailed knowledge may occasionally enable one to say ‘The pauperism is low here, since the proportion of out-relief is very small,’ or perhaps, ‘The proportion of out-relief given is large on account of the high pauperism and other industrial conditions of the union’: but such cases will be exceptional and will as a rule only refer to large deviations from the mean.” Yule, “On the Correlation of Total Pauperism with Proportion of Out-Relief,” 1895, 605n2; discussed in Mills, Statistical Biography, 46.
19. Yule，《对最近两个人口普查间隔年内英国贫困化变化原因的调查（第一部分）》，第251页。
19. Yule, “An Investigation into the Causes of Changes in Pauperism in England, Chiefly During the Last Two Intercensal Decades (Part I.),” 251.
20. “根据与贫困和救济比例相关的因素被忽略的数量，仍然有一定的误差可能性，但显然这种误差的可能性比以前小得多。” Yule，251。
20. “There is still a certain chance of error depending on the number of factors correlated both with pauperism and with proportion of out-relief which have been omitted, but obviously this chance of error will be much smaller than before.” Yule, 251.
21. 斯蒂格勒，《统计学史》，356。
21. Stigler, The History of Statistics, 356.
22. Yule，《对最近两个人口普查间隔年内英国贫困化变化原因的调查（第一部分）》，第265页。
22. Yule, “An Investigation into the Causes of Changes in Pauperism in England, Chiefly During the Last Two Intercensal Decades (Part I.),” 265.
23. Yule，257n.16。
23. Yule, 257n.16.
24. 圣诞节，277。
24. Yule, 277.
25. Yule，《论相关理论》，第812页。
25. Yule, “On the Theory of Correlation,” 812.
26. Arthur Cecil Pigou，《关于济贫法救济的一些经济方面和影响的备忘录》，载《济贫法和救济困境皇家委员会》，附录，第 9 卷，《1910 年 2 月 15 日至 11 月 28 日会议的议会文件》，第 49 卷。（伦敦：陛下文书局，1910 年），第 984-85 页。
26. Arthur Cecil Pigou, “Memorandum on Some Economic Aspects and Effects of Poor Law Relief,” in Royal Commission on the Poor Laws and Relief of Distress, Appendix, Vol. 9, Parliamentary Papers for the Session 15 February 1910–28 November 1910, Vol. 49. (London: His Majesty’s Stationery Office, 1910), 984–85.
27. 庇古，986。
27. Pigou, 986.
28. 庇古，986。
28. Pigou, 986.
29. David A. Freedman，“统计模型和鞋革”，社会学方法论21 (1991): 291，https://doi.org/10.2307/270939。
29. David A. Freedman, “Statistical Models and Shoe Leather,” Sociological Methodology 21 (1991): 291, https://doi.org/10.2307/270939.
30. Yule，《对最近两个人口普查间隔年的英国贫困化变化原因的调查（第一部分）》，第270页。
30. Yule, “An Investigation into the Causes of Changes in Pauperism in England, Chiefly During the Last Two Intercensal Decades (Part I.),” 270.
31.Desrosières ，《大数政治》，140。
31. Desrosières, The Politics of Large Numbers, 140.
32. Shivrang Setlur，“寻找南亚情报：1919-1940 年英属印度的心理测量学”，《行为科学史杂志》 50 卷，第 4 期（2014 年）：359-75 页，https://doi.org/10.1002/jhbs.21692。
32. Shivrang Setlur, “Searching for South Asian Intelligence: Psychometry in British India, 1919–1940,” Journal of the History of the Behavioral Sciences 50, no. 4 (2014): 359–75, https://doi.org/10.1002/jhbs.21692.
33. Charles Spearman，“‘一般智力’的客观确定和测量”，《美国心理学杂志》第 15 卷，第 2 期（1904 年 4 月）：277，https://doi.org/10.2307/1412107（斜体为本刊所加）。
33. Charles Spearman, “ ‘General Intelligence,’ Objectively Determined and Measured,” The American Journal of Psychology 15, no. 2 (April 1904): 277, https://doi.org/10.2307/1412107 (our italics).
34. Adrian Wooldridge，《测量心智：1860-1990年左右的英国教育与心理学》（纽约：剑桥大学出版社，1994年），第74页。
34. Adrian Wooldridge, Measuring the Mind: Education and Psychology in England, c. 1860–c. 1990 (New York: Cambridge University Press, 1994), 74.
35. 查尔斯·斯皮尔曼，《“智力”的本质与认知原理》（伦敦：麦克米伦，1923年），第355页；引自史蒂芬·杰·古尔德，《人的错误测量》（纽约：诺顿，1996年），第293页。
35. Charles Spearman, The Nature of “Intelligence” and the Principles of Cognition (London: Macmillan, 1923), 355; quoted in Stephen Jay Gould, The Mismeasure of Man (New York: Norton, 1996), 293.
36. 查尔斯·斯皮尔曼，《人的能力：其性质与衡量标准》（纽约：麦克米伦公司，1927年），第379页；转引自古尔德，《人的错误衡量》，第301、302页。
36. Charles Spearman, The Abilities of Man: Their Nature and Measurement (New York: The Macmillan Company, 1927), 379; quoted in Gould, The Mismeasure of Man, 301, 302.
37.Spearman ，《人类的能力；其性质及其衡量》，380。
37. Spearman, The Abilities of Man; Their Nature and Measurement, 380.
38. 约翰·卡森（John Carson）的《功绩的衡量标准：1750-1940 年法兰西共和国和美利坚共和国的才能、智力和不平等》（新泽西州普林斯顿：普林斯顿大学出版社，2007 年），第 183-193 页对有关衡量智力的争论进行了出色的调查。
38. John Carson, The Measure of Merit: Talents, Intelligence, and Inequality in the French and American Republics, 1750–1940 (Princeton, NJ: Princeton University Press, 2007), 183–93 provides an excellent survey of debates about measuring intelligence.
39. Karl Pearson 和 Margaret Moul，《智力的数学。广义因子理论中的抽样误差》， Biometrika 19, no. 3/4 (1927): 291，https://doi.org/10.2307/2331962。请参阅 Theodore M. Porter 的《Karl Pearson：统计时代的科学生活》（新泽西州普林斯顿：普林斯顿大学出版社，2004 年），第 270 页。
39. Karl Pearson and Margaret Moul, “The Mathematics of Intelligence. The Sampling Errors in the Theory of a Generalised Factor,” Biometrika 19, no. 3/4 (1927): 291, https://doi.org/10.2307/2331962. See Theodore M. Porter, Karl Pearson: The Scientific Life in a Statistical Age (Princeton, NJ: Princeton University Press, 2004), 270.
40. 卡森，《功绩的衡量》，159。
40. Carson, The Measure of Merit, 159.
41. Richard J. Herrnstein 和 Charles A. Murray，《钟形曲线：美国生活中的智力和阶级结构》（纽约：西蒙与舒斯特出版社，1996 年）。
41. Richard J. Herrnstein and Charles A. Murray, The Bell Curve: Intelligence and Class Structure in American Life (New York: Simon & Schuster, 1996).
42. Colin Koopman，《我们如何成为我们的数据：信息人的谱系》（芝加哥：芝加哥大学出版社，2019 年）。19 世纪，被奴役的人受到了仔细的核算，参见 Caitlin Rosenthal，《奴隶制的核算：主人与管理》（马萨诸塞州剑桥：哈佛大学出版社，2018 年），尤其是第 2 章；Simone Browne，《黑暗物质：关于对黑人的监视》（北卡罗来纳州达勒姆：杜克大学出版社，2015 年）。
42. Colin Koopman, How We Became Our Data: A Genealogy of the Informational Person (Chicago: The University of Chicago Press, 2019). Enslaved people had been subject to careful accounting in the nineteenth century, see Caitlin Rosenthal, Accounting for Slavery: Masters and Management (Cambridge, MA: Harvard University Press, 2018), esp. ch. 2; Simone Browne, Dark Matters: On the Surveillance of Blackness (Durham, NC: Duke University Press, 2015).
43. Wangui Muigai，Projit Bihari Mukharji 等人，“关于收集人口统计数据的圆桌讨论”，Isis 111，第 1 期。 2（2020 年 6 月）：320，https://doi.org/10.1086/709484。
43. Wangui Muigai, in Projit Bihari Mukharji et al., “A Roundtable Discussion on Collecting Demographics Data,” Isis 111, no. 2 (June 2020): 320, https://doi.org/10.1086/709484.
44. Wangui Muigai，载于 Mukharji 等人，第 320 页。有关性别类别假设的持续显著性，请参阅 Mar Hicks，“Hacking the Cis-Tem”，IEEE 计算史年鉴41，第 1 期（2019 年 1 月）：20–33，https://doi.org/10.1109/MAHC.2019.2897667。
44. Wangui Muigai, in Mukharji et al., 320. For the continuing salience of assumptions of gender categories, see Mar Hicks, “Hacking the Cis-Tem,” IEEE Annals of the History of Computing 41, no. 1 (January 2019): 20–33, https://doi.org/10.1109/MAHC.2019.2897667.
45. 引用自 Sandeep Mertia，《马哈拉诺比斯梦见过机器人吗？》，《数据生活：印度计算文化论文集》，Sandeep Mertia 和 Ravi Sundaram 编辑（阿姆斯特丹：网络文化研究所，2020 年），第 31 页。
45. Quoted in Sandeep Mertia, “Did Mahalanobis Dream of Androids?,” in Lives of Data: Essays on Computational Cultures from India, ed. Sandeep Mertia and Ravi Sundaram (Amsterdam: Institute of Network Cultures, 2020), 31.
46. 梅蒂亚，31。
46. Mertia, 31.
47. Emmanuel Didier，《数字中的美国：量化、民主和国家统计的诞生》（马萨诸塞州剑桥：麻省理工学院出版社，2020 年），第 11 页。
47. Emmanuel Didier, America by the Numbers: Quantification, Democracy, and the Birth of National Statistics (Cambridge, MA: MIT Press, 2020), 11.
48. J. Adam Tooze，《1900-1945年的统计数据与德国国家：现代经济知识的形成》（纽约：剑桥大学出版社，2001年），第24页
48. J. Adam Tooze, Statistics and the German State, 1900–1945: The Making of Modern Economic Knowledge (New York: Cambridge University Press, 2001), 24
49. Arunabh Ghosh，《让统计发挥作用：中华人民共和国早期的统计与治国方略》（普林斯顿：普林斯顿大学出版社，2020 年），第 283 页。
49. Arunabh Ghosh, Making It Count: Statistics and Statecraft in the Early People’s Republic of China (Princeton: Princeton University Press, 2020), 283.
50. Tooze，《统计学与德国国家》，28页。
50. Tooze, Statistics and the German State, 28.
51. John Koren 和 Edmund Ezra Day，《统计学的历史及其在许多国家的发展和进步》，载《纪念美国统计协会成立七十五周年回忆录》（纽约：麦克米伦公司为美国统计协会出版，1918 年），第 25-26 页，http://archive.org/details/cu31924013894997。
51. John Koren and Edmund Ezra Day, The History of Statistics, Their Development and Progress in Many Countries; in Memoirs to Commemorate the Seventy Fifth Anniversary of the American Statistical Association (New York: Pub. for the American Statistical Association by the Macmillan Company of New York, 1918), 25–26, http://archive.org/details/ cu31924013894997.
52. 约翰·斯图尔特·密尔，《政治经济学原理：及其在社会哲学中的一些应用》（伦敦：JW Parker，1848 年），第375页。
52. John Stuart Mill, Principles of Political Economy: With Some of Their Applications to Social Philosophy (London: J. W. Parker, 1848), 375.
53. Kevin Bird，“仍然不在我们的基因中：抵制围绕 GWAS 的叙述”，《科学为民》杂志第 23 卷，第 3 期（2021 年 2 月 5 日），https://magazine.scienceforthepeople.org/vol23–3-bio-politics/genetic-basis-genome-wide-association-studies-risk/。
53. Kevin Bird, “Still Not in Our Genes: Resisting the Narrative Around GWAS,” Science for the People Magazine 23, no. 3 (February 5, 2021), https://magazine.scienceforthepeople.org/vol23–3-bio-politics/genetic-basis -genome-wide-association-studies-risk/.
54. Kelly Miller，《霍夫曼的《美国黑人的种族特征和倾向》评论》，《美国黑人学院》。临时论文，第 1 期（华盛顿特区：学院，1897 年），第 35 页，https://catalog.hathitrust.org/Record/100788175。
54. Kelly Miller, A Review of Hoffman’s Race Traits and Tendencies of the American Negro, American Negro Academy. Occasional Papers, no. 1 (Washington, DC: The Academy, 1897), 35, https://catalog.hathitrust.org/ Record/100788175.
第五章 数据的数学洗礼
CHAPTER 5: DATA’S MATHEMATICAL BAPTISM
1. Joan Fisher Box，《Guinness、Gosset、Fisher 和小样本》，《统计科学》 2 卷，第 1 期（1987 年）：48。
1. Joan Fisher Box, “Guinness, Gosset, Fisher, and Small Samples,” Statistical Science 2, no. 1 (1987): 48.
2.学生，“论谷物品种的测试”，Biometrika 15，第 3/4 期（1923 年）：271，https://doi.org/10.2307/2331868。
2. Student, “On Testing Varieties of Cereals,” Biometrika 15, no. 3/4 (1923): 271, https://doi.org/10.2307/2331868.
3.学生，“平均值的可能误差”，Biometrika 6，第 1 期（1908）：2，https://doi.org/10.2307/2331554。
3. Student, “The Probable Error of a Mean,” Biometrika 6, no. 1 (1908): 2, https://doi.org/10.2307/2331554.
4. Box，《吉尼斯、戈塞特、费舍尔和小样本》。Box 完美地唤起了人们对戈塞特的世界和作品的回忆。
4. Box, “Guinness, Gosset, Fisher, and Small Samples.” Box wonderfully evokes Gosset’s world and work.
5. ES Pearson，“‘学生’作为统计学家”，Biometrika 30，第 3/4 期（1939 年 1 月）：215–16，https://doi.org/10.2307/2332648。
5. E. S. Pearson, “ ‘Student’ as Statistician,” Biometrika 30, no. 3/4 (January 1939): 215–16, https://doi.org/10.2307/2332648.
6. Box，《Guinness、Gosset、Fisher 和小样本》，第 49 页。
6. Box, “Guinness, Gosset, Fisher, and Small Samples,” 49.
7. Donald A. MacKenzie，《1865-1930 年英国的统计学：科学知识的社会建构》（爱丁堡：爱丁堡大学出版社，1981 年），111f 页；若想更深入地了解 Gosset 的作品，请特别参阅 Stephen Thomas Ziliak 和 Deirdre N. McCloskey 的《统计显著性崇拜：标准误差如何让我们失去工作、公正和生命》（安娜堡：密歇根大学出版社，2008 年）。
7. Donald A. MacKenzie, Statistics in Britain, 1865–1930: The Social Construction of Scientific Knowledge (Edinburgh: Edinburgh University Press, 1981), 111f; for a readable appreciation of Gosset, see especially Stephen Thomas Ziliak and Deirdre N. McCloskey, The Cult of Statistical Significance: How the Standard Error Costs Us Jobs, Justice, and Lives (Ann Arbor: University of Michigan Press, 2008).
8. Nan M. Laird，“与 FN David 的对话”，《统计科学》 4，第 3 期（1989 年 8 月）：238，https://doi.org/10.1214/ss/1177012487。
8. Nan M. Laird, “A Conversation with F. N. David,” Statistical Science 4, no. 3 (August 1989): 238, https://doi.org/10.1214/ss/1177012487.
9.Ronald Aylmer Fisher，《研究人员的统计方法》（伦敦：Oliver and Boyd，1925 年），vii。
9. Ronald Aylmer Fisher, Statistical Methods for Research Workers (London: Oliver and Boyd, 1925), vii.
10. 请参阅 Giuditta Parolini，“现代统计学在农业科学中的兴起：方差分析、实验设计和罗瑟姆斯特德实验站研究的重塑，1919-1933”，《生物学史杂志》第 48 卷，第 2 期（2015 年 5 月）：301-35，https://doi.org/10.1007/s10739-014-9394-z。
10. See especially Giuditta Parolini, “The Emergence of Modern Statistics in Agricultural Science: Analysis of Variance, Experimental Design and the Reshaping of Research at Rothamsted Experimental Station, 1919–1933,” Journal of the History of Biology 48, no. 2 (May 2015): 301–35, https://doi .org/10.1007/s10739–014–9394-z.
11. Box，“Guinness、Gosset、Fisher 和小样本”，第 51 页。
11. Box, “Guinness, Gosset, Fisher, and Small Samples,” 51.
12. EL Lehmann，《Fisher、Neyman 和古典统计学的创立》（纽约：Springer，2011 年），第 12 页。
12. E. L. Lehmann, Fisher, Neyman, and the Creation of Classical Statistics (New York: Springer, 2011), 12.
13. RA Fisher 和 WA Mackenzie，《作物变异研究。II.不同马铃薯品种对粪肥的反应》，《农业科学杂志》 13（1923 年）：469。
13. R. A. Fisher and W. A. Mackenzie, “Studies in Crop Variation. II. The Manurial Response of Different Potato Varieties,” The Journal of Agricultural Science 13 (1923): 469.
14. Fisher，《研究人员的统计方法》，vii。
14. Fisher, Statistical Methods for Research Workers, vii.
15. 费舍尔，4。
15. Fisher, 4.
16.Ronald Aylmer Fisher，《实验设计》（伦敦：Oliver and Boyd，1935 年），第 15-16 页。
16. Ronald Aylmer Fisher, The Design of Experiments (London: Oliver and Boyd, 1935), 15–16.
17. Fisher, 49。有关随机化的发展，请参阅 Nancy S. Hall，“RA Fisher 及其对随机化的倡导”，《生物学史杂志》 40，第 2 期（2007 年 6 月 1 日）：295–325，https://doi.org/10.1007/s10739–006–9119-z。
17. Fisher, 49. For the development of randomization, see Nancy S. Hall, “R. A. Fisher and His Advocacy of Randomization,” Journal of the History of Biology 40, no. 2 (June 1, 2007): 295–325, https://doi.org/10.1007/s10739–006–9119-z.
18. Epstein，《不纯的科学：艾滋病、激进主义和知识政治》（伯克利：加州大学出版社，1996 年）。
18. Epstein, Impure Science: AIDS, Activism, and the Politics of Knowledge (Berkeley: University of California Press, 1996).
19. Stephen T. Ziliak，“WS Gosset 和实验统计学中一些被忽视的概念：Guinnessometrics II*”，《葡萄酒经济学杂志》第 6 卷，第 2 期（2011 年编辑）：252–77，https://doi.org/10.1017/S1931436100001632。
19. Stephen T. Ziliak, “W.S. Gosset and Some Neglected Concepts in Experimental Statistics: Guinnessometrics II*,” Journal of Wine Economics 6, no. 2 (ed 2011): 252–77, https://doi.org/10.1017/S1931436100001632.
20.Fisher ，《实验设计》，第10页。
20. Fisher, The Design of Experiments, 10.
21. 费舍尔，10。
21. Fisher, 10.
22. Ronald Aylmer Fisher，《优生学家的一些希望》， 《R.A . Fisher文集》第1卷（阿德莱德：阿德莱德大学，1971年），78. 有关费舍尔作为优生学家的研究，请参阅 Alex Aylward，《RA Fisher、优生学和两次世界大战期间英国的家庭津贴运动》，《英国科学史杂志》第 54 卷，第 4 期（2021 年 12 月）：485–505，https://doi.org/10.1017/S0007087421000674。
22. Ronald Aylmer Fisher, “Some Hopes of a Eugenicist,” in Collected Papers of R. A. Fisher, vol. 1 (Adelaide: University of Adelaide, 1971), 78. For Fisher as a eugenicist, see Alex Aylward, “R.A. Fisher, Eugenics, and the Campaign for Family Allowances in Interwar Britain,” The British Journal for the History of Science 54, no. 4 (December 2021): 485–505, https://doi .org/10.1017/S0007087421000674.
23. Fisher，《优生学家的一些希望》，第79页。
23. Fisher, “Some Hopes of a Eugenicist,” 79.
24. Ronald Fisher，“统计方法和科学归纳”，皇家统计学会杂志：B 系列（方法论） 17，第 1 期（1955 年 1 月 1 日）：75，https://doi.org/10.1111/j.2517–6161.1955.tb00180.x。
24. Ronald Fisher, “Statistical Methods and Scientific Induction,” Journal of the Royal Statistical Society: Series B (Methodological) 17, no. 1 (January 1, 1955): 75, https://doi.org/10.1111/j.2517–6161.1955.tb00180.x.
25. J. Neyman 和 ES Pearson，《论统计假设最有效检验问题》，《伦敦皇家学会哲学学报》，A辑，包含数学或物理论文，第 231 卷（1933 年）：第 291 页。
25. J. Neyman and E. S. Pearson, “On the Problem of the Most Efficient Tests of Statistical Hypotheses,” Philosophical Transactions of the Royal Society of London. Series A, Containing Papers of a Mathematical or Physical Character 231 (1933): 291.
26. Theodora Dryer，“设计确定性：1920-1970 年焦虑时代算法计算的兴起”（博士论文，加州大学圣地亚哥分校，2019 年），第 81 页。
26. Theodora Dryer, “Designing Certainty: The Rise of Algorithmic Computing in an Age of Anxiety 1920–1970” (PhD Thesis, UC San Diego, 2019), 81.
27. Constance Reid，Neyman（纽约：Springer，1998），24–25，48。她描述了一份手稿，其第一部分“涉及证明在自然现象研究中使用抽象数学理论的合理性的原理，特别是在农业实验领域。”
27. Constance Reid, Neyman (New York: Springer, 1998), 24–25, 48. She’s describing a manuscript whose first part “is concerned with principles which justify the use of abstract mathematical theory in studies of natural phenomena, especially in the domain of agricultural experimentation.”
28. J. Neyman，“‘归纳行为’作为科学哲学的基本概念”，《国际统计研究所评论》第 25 卷，第 1/3 期（1957 年）：8 页，https://doi.org/10.2307/1401671。
28. J. Neyman, “ ‘Inductive Behavior’ as a Basic Concept of Philosophy of Science,” Revue de l’Institut International de Statistique / Review of the International Statistical Institute 25, no. 1/3 (1957): 8, https://doi.org/10 .2307/1401671.
29. Karl Pearson，《科学语法》（伦敦：Walter Scott；纽约：Charles Scribner's Sons，1892 年），72，http://archive.org/details/grammar ofscience00pearrich。
29. Karl Pearson, The Grammar of Science (London: Walter Scott ; New York: Charles Scribner’s Sons, 1892), 72, http://archive.org/details/grammar ofscience00pearrich.
30. Gosset 致 Egon Pearson，1926 年 5 月 11 日，收录于 Pearson，《‘学生’作为统计学家》，第 243 页。另请参阅 Lehmann、Fisher、Neyman 和古典统计学的创立，第 7 页。
30. Gosset to Egon Pearson, 11.5.1926, in Pearson, “ ‘Student’ as Statistician,” 243. See Lehmann, Fisher, Neyman, and the Creation of Classical Statistics, 7.
31. Gosset 致 Egon Pearson，1926 年 5 月 11 日，载于 Pearson，《‘学生’作为统计学家》，第 242 页。
31. Gosset to Egon Pearson, 11.5.1926, in Pearson, “ ‘Student’ as Statistician,” 242.
32. Jerzy Neyman，《J. Neyman 早期统计论文选集》（伯克利：加州大学出版社，1967 年），第 352 页。
32. Jerzy Neyman, A Selection of Early Statistical Papers of J. Neyman. (Berkeley: University of California Press, 1967), 352.
33. Lehmann、Fisher、Neyman 和古典统计学的创立，第 37 页（斜体为我方所加）。
33. Lehmann, Fisher, Neyman, and the Creation of Classical Statistics, 37 (our italics).
34. Lehmann、Fisher、Neyman 和古典统计学的创立； Gerd Gigerenzer 编，《机会帝国：概率如何改变科学和日常生活》（英国剑桥：剑桥大学出版社，1989 年）。有关社会科学的后续发展，请参阅 Hunter Heyck，《系统时代》（巴尔的摩：约翰霍普金斯大学出版社，2015 年），第 4 章。
34. Lehmann, Fisher, Neyman, and the Creation of Classical Statistics; Gerd Gigerenzer, ed., The Empire of Chance: How Probability Changed Science and Everyday Life (Cambridge, UK: Cambridge University Press, 1989). For the subsequent big picture in the social sciences, see Hunter Heyck, Age of System (Baltimore: Johns Hopkins University Press, 2015), ch. 4.
35.Neyman ，《J.Neyman早期统计论文选集》，352。
35. Neyman, A Selection of Early Statistical Papers of J. Neyman, 352.
36. Ronald Aylmer Fisher，《科学思想与人类推理的精炼》，《日本运筹学会刊》第 3 期（1960 年）：3。Justin Joque，《革命性的数学：人工智能、统计学和资本主义的逻辑》（纽约：Verso，2022 年），第 5 章对这场竞赛进行了丰富的哲学解读。
36. Ronald Aylmer Fisher, “Scientific Thought and the Refinement of Human Reasoning,” Journal of the Operations Research Society of Japan 3 (1960): 3. Justin Joque, Revolutionary Mathematics: Artificial Intelligence, Statistics and the Logic of Capitalism (New York: Verso, 2022), ch. 5 offers a rich philosophical reading of this contest.
37. Ronald Aylmer Fisher，《统计方法和科学推断》（爱丁堡：Oliver and Boyd，1956 年），第 7 页，http://archive.org/details/statisticalmetho0000fish。
37. Ronald Aylmer Fisher, Statistical Methods and Scientific Inference (Edinburgh: Oliver and Boyd, 1956), 7, http://archive.org/details/statisticalmetho 0000fish.
38. Gerd Gigerenzer 和 Julian N. Marewski，“替代科学：科学推理的通用方法的偶像”，《管理学期刊》第 41 卷第 2 期（2015 年 2 月 1 日）：第 421–40 页，https://doi.org/10.1177/0149206314547522。
38. Gerd Gigerenzer and Julian N. Marewski, “Surrogate Science: The Idol of a Universal Method for Scientific Inference,” Journal of Management 41, no. 2 (February 1, 2015): 421–40, https://doi.org/10.1177/0149206314547522.
39. Snedecor 的统计方法是一个关键载体。
39. Snedecor’s Statistical Methods was a key vector.
40. Christopher Phillips，《推理仪式：算法和统计史》，《算法现代性：机械化思想和行动，1500-2000》，Massimo Mazzotti 和 Morgan Ames 编（英国牛津：牛津大学出版社，即将出版）。
40. Christopher Phillips, “Inference Rituals: Algorithms and the History of Statistics,” in Algorithmic Modernity: Mechanizing Thought and Action, 1500–2000, ed. Massimo Mazzotti and Morgan Ames (Oxford, UK: Oxford University Press, forthcoming).
41. Theodore M. Porter，《信任数字：追求科学与公共生活中的客观性》（新泽西州普林斯顿：普林斯顿大学出版社，1995年），第206页。
41. Theodore M. Porter, Trust in Numbers: The Pursuit of Objectivity in Science and Public Life (Princeton, NJ: Princeton University Press, 1995), 206.
42. 波特，206。
42. Porter, 206.
43. Gigerenzer，《机会帝国》，106。
43. Gigerenzer, The Empire of Chance, 106.
44. W. Allen Wallis，“统计研究小组，1942-1945”，《美国统计协会杂志》 75，第 370 期（1980 年 6 月 1 日）：321，https://doi.org/10.1080/01621459.1980.10477469。
44. W. Allen Wallis, “The Statistical Research Group, 1942–1945,” Journal of the American Statistical Association 75, no. 370 (June 1, 1980): 321, https:// doi.org/10.1080/01621459.1980.10477469.
45. Judy L. Klein，“为客户讲经济学：统计质量控制和序贯分析案例”，《政治经济学史》第 32 卷，第 1 期增刊（2000 年）：第 25-70 页；Nicola Giocoli，“从瓦尔德到野蛮人：经济人成为贝叶斯统计学家”，《行为科学史杂志》第 49 卷，第 1 期（2013 年）：第 63-95 页，https://doi.org/10.1002/jhbs.21579。
45. Judy L. Klein, “Economics for a Client: The Case of Statistical Quality Control and Sequential Analysis,” History of Political Economy 32, no. Suppl. 1 (2000): 25–70; Nicola Giocoli, “From Wald to Savage: Homo Economicus Becomes a Bayesian Statistician,” Journal of the History of the Behavioral Sciences 49, no. 1 (2013): 63–95, https://doi.org/10.1002/jhbs.21579.
46. [Mina Rees]，ONR 数学项目描述，1946 年 9 月 27 日，Hotelling 论文，第 18 盒，ONR 合同和续约，哥伦比亚大学特藏。
46. [Mina Rees], description of mathematics program of ONR, 9/27/1946, Hotelling Papers, Box 18, ONR Contract and Renewals, Columbia University Special Collections.
47. Harold Hotelling，“统计学在大学中的地位（附讨论）”，载《第一届伯克利数理统计和概率研讨会论文集》（加利福尼亚州伯克利，加利福尼亚大学董事会，1949 年），第 23 页，https://projecteuclid.org/euclid.bsmsp/1166219196。
47. Harold Hotelling, “The Place of Statistics in the University (with Discussion),” in Proceedings of the [First] Berkeley Symposium on Mathematical Statistics and Probability (Berkeley, CA, The Regents of the University of California, 1949), 23, https://projecteuclid.org/euclid.bsmsp/1166219196.
48. Jerzy Neyman 编辑，《第一届伯克利数理统计和概率研讨会论文集》（加利福尼亚州伯克利：加利福尼亚大学董事会，1949 年），https://projecteuclid.org/euclid.bsmsp /1166219194。
48. Jerzy Neyman, ed., Proceedings of the [First] Berkeley Symposium on Mathematical Statistics and Probability (Berkeley, CA: The Regents of the University of California, 1949), https://projecteuclid.org/euclid.bsmsp /1166219194.
49. Hotelling，“统计学在大学中的地位（附讨论）”，第 23 页。
49. Hotelling, “The Place of Statistics in the University (with Discussion),” 23.
50. John W. Tukey，《数据分析的未来》，《数理统计年鉴》第 33 卷，第 1 期（1962 年）：第 6 页。
50. John W. Tukey, “The Future of Data Analysis,” The Annals of Mathematical Statistics 33, no. 1 (1962): 6.
第六章：战争数据
CHAPTER 6: DATA AT WAR
1. Juanita Moody，《口述历史》，Jean Lichty 等人采访，1994 年 6 月 16 日，26，https://media.def ense.gov/2021/Jul/15/2002763502/-1/-1/0/NSA-OH -1994–32-MOODY.PDF。
1. Juanita Moody, Oral History, interview by Jean Lichty et al., June 16, 1994, 26, https://media.def ense.gov/2021/Jul/15/2002763502/-1/-1/0/NSA-OH -1994–32-MOODY.PDF.
2.马克·布朗，“布莱切利披露 1938 年‘射击派对’的真实意图” ， 《卫报》，2018 年 9 月 18 日，sec. 世界新闻，https://www.theguardian.com/world/2018/sep/18/bletchley-discloses-real-intention-1938-shooting-party-wapark-r。
2. Mark Brown, “Bletchley Discloses Real Intention of 1938 ‘Shooting Party,’ ” The Guardian, September 18, 2018, sec. World news, https://www.theguardian .com/world/2018/sep/18/bletchley-discloses-real-intention-1938-shooting-party -wapark-r.
3. Howard Campaigne，《口述历史》，罗伯特·D·法利 (Robert D Farley) 访谈，1983 年 6 月 29 日，第 15-16 页，https://www.nsa.gov/portals/75/documents/news-features/declassified-documents/oral-history-interviews/nsa-oh-14-83-campaigne.pdf。
3. Howard Campaigne, Oral History, interview by Robert D Farley, June 29, 1983, 15–16, https://www.nsa.gov/portals/75/documents/news -features/declassified-documents/oral-history-interviews/nsa-oh-14–83 -campaigne.pdf.
4. David Kenyon，《布莱切利园与 D 日》（康涅狄格州纽黑文：耶鲁大学出版社，2019 年），第 236 页。
4. David Kenyon, Bletchley Park and D-Day (New Haven, CT: Yale University Press, 2019), 236.
5.埃莉诺·爱尔兰 (Eleanor Ireland)，《口述历史》，珍妮特·阿巴特 (Janet Abbate) 采访，2001 年 4 月 23 日，https://ethw.org/Oral-History:Eleanor_Ireland。
5. Eleanor Ireland, Oral History, interview by Janet Abbate, April 23, 2001, https://ethw.org/Oral-History:Eleanor_Ireland.
6. J. Abbate，《重新编码性别：女性在计算领域的参与度变化》（马萨诸塞州剑桥：麻省理工学院出版社，2012 年），第 20 页。另请参阅 Mar Hicks， 《程序化的不平等：英国如何抛弃女性技术人员并失去计算优势》（马萨诸塞州剑桥：麻省理工学院出版社，2017 年），第 1 章。
6. J. Abbate, Recoding Gender: Women’s Changing Participation in Computing (Cambridge, MA: MIT Press, 2012), 20. See also Mar Hicks, Programmed Inequality: How Britain Discarded Women Technologists and Lost Its Edge in Computing (Cambridge, MA: MIT Press, 2017), chap. 1.
7. Abbate，《重新编码性别：女性在计算领域的参与变化》，22。
7. Abbate, Recoding Gender: Women’s Changing Participation in Computing, 22.
8.阿巴特 (Abbate)，27 岁，根据她在爱尔兰的采访，撰写了《口述历史》。
8. Abbate, 27. Drawing upon her interview with Ireland, Oral History.
9.转引自 B. Jack Copeland 主编的《巨人：布莱切利园密码破译计算机的秘密》（牛津；纽约：牛津大学出版社，2006 年），第 171 页。
9. Quoted in B. Jack Copeland, ed., Colossus: The Secrets of Bletchley Park’s Codebreaking Computers (Oxford; New York: Oxford University Press, 2006), 171.
10. 希克斯，《程序化不平等》，第 40-41 页。
10. Hicks, Programmed Inequality, 40–41.
11. Abraham Sinkov，《口述历史》，Arthur J Zoebelein 等人采访，1979 年 5 月，第 3-4 页。
11. Abraham Sinkov, Oral History, interview by Arthur J Zoebelein et al., May 1979, 3–4.
12. 所罗门·库尔贝克，《口述历史》，RD Farley 和 HF Schorreck 采访，1982 年 8 月 26 日，第 48 页。
12. Solomon Kullback, Oral History, interview by R. D. Farley and H. F. Schorreck, August 26, 1982, 48.
13. Phillip Rogaway，“密码工作的道德品质”（2015），1，https://web.cs.ucdavis.edu/~rogaway/papers/moral-fn.pdf。
13. Phillip Rogaway, “The Moral Character of Cryptographic Work” (2015), 1, https://web.cs.ucdavis.edu/~rogaway/papers/moral-fn.pdf.
14. WJ Holmes，《双刃秘密：二战期间美国海军在太平洋的情报行动》（马里兰州安纳波利斯：海军学院出版社，2012 年），第 142 页。
14. W. J. Holmes, Double Edged Secrets: U.S. Naval Intelligence Operations in the Pacific During World War II. (Annapolis, MD: Naval Institute Press, 2012), p. 142.
15. Kenyon，《布莱切利园与诺曼底登陆》，第 242–43 页。
15. Kenyon, Bletchley Park and D-Day, 242–43.
16. 第一个假设对费舍尔来说相当于零假设，或者对奈曼及其学派来说，它既相当于零假设，又相当于竞争假设。
16. The first is equivalent to the null hypothesis for Fisher, or to both the null hypothesis and the competing hypothesis for Neyman and his school.
17. Stephen M. Stigler，“贝叶斯论文的真实标题”，《统计科学》第 28 卷，第 3 期（2013 年 8 月）：283–88 页，https://doi.org/10.1214/13-STS438；Richard Swinburne，“贝叶斯、上帝和多元宇宙”，《宗教哲学中的概率》，Jake Chandler 和 Victoria S. Harrison 主编（英国牛津：牛津大学出版社，2012 年），103–26 页。
17. Stephen M. Stigler, “The True Title of Bayes’s Essay,” Statistical Science 28, no. 3 (August 2013): 283–88, https://doi.org/10.1214/13-STS438; Richard Swinburne, “Bayes, God, and the Multiverse,” in Probability in the Philosophy of Religion, ed. Jake Chandler and Victoria S. Harrison (Oxford, UK: Oxford University Press, 2012), 103–26.
18. 此外，我们可能对于应该考虑多少个假设存在分歧。
18. Moreover, we may disagree on how many hypotheses should be considered.
19. Ian Taylor，“Alan M. Turing：概率在密码学中的应用”，ArXiv:1505.04714 [数学]，2015 年 5 月 26 日，第 3 页，http://arxiv.org/abs/1505.04714。
19. Ian Taylor, “Alan M. Turing: The Applications of Probability to Cryptography,” ArXiv:1505.04714 [Math], May 26, 2015, 3, http://arxiv.org/abs/1505 .04714.
20. 有关图灵的方法，请参阅 Sandy Zabell，“关于艾伦·M·图灵的评论：概率在密码学中的应用”，Cryptologia 36，第 3 期（2012 年 7 月）：191–214，https://doi.org/10.1080/01611194.2012.697811。
20. For Turing’s approach, see Sandy Zabell, “Commentary on Alan M. Turing: The Applications of Probability to Cryptography,” Cryptologia 36, no. 3 (July 2012): 191–214, https://doi.org/10.1080/01611194.2012.697811.
21. FT Leahy，“贝叶斯因子的明显悖论（U）”，NSA 技术期刊27，第 3 期（nd）：8，9。比较图灵本人的评论，Taylor，“Alan M. Turing”，2-3。
21. F. T. Leahy, “The Apparent Paradox of Bayes Factors (U),” NSA Technical Journal 27, no. 3 (n.d.): 8, 9. Compare Turing’s own remarks, Taylor, “Alan M. Turing,” 2–3.
22. 有关贝叶斯及其成功的通俗历史，请参阅 SB McGrayne 的《永不消亡的理论：贝叶斯规则如何破解恩尼格玛密码、追捕俄罗斯潜艇并在两个世纪的争议中取得胜利》（康涅狄格州纽黑文：耶鲁大学出版社，2011 年）。有关哲学性更强的读物，请参阅 Joque 的《革命数学》，第 6-7 章。
22. For a popular history of Bayes and its successes, see S. B. McGrayne, The Theory That Would Not Die: How Bayes’ Rule Cracked the Enigma Code, Hunted Down Russian Submarines, & Emerged Triumphant from Two Centuries of Controversy (New Haven, CT: Yale University Press, 2011). And for a more philosophical reading, Joque, Revolutionary Mathematics, chs. 6 -7.
23. 在文件解密后，Good 明确地写到他在战争期间工作的根源。Irving J. Good，《图灵对经验贝叶斯的预测与海军密码分析》，《统计计算与模拟杂志》第 66 卷，第 2 期（2000 年）：101-11。他的作品也出现在美国国家安全局的机密杂志中：Irving J. Good，《贝叶斯-图灵因子的属性列表》，《美国国家安全局技术杂志》第 10 卷，第 2 期（1965 年），https://www.nsa.gov/Portals/70/documents/news-features/declassified-documents/tech-journals/list-of-properties.pdf。
23. As documents were declassified, Good wrote explicitly about the roots of his work during the war. Irving J. Good, “Turing’s Anticipation of Empirical Bayes in Connection with the Cryptanalysis of the Naval Enigma,” Journal of Statistical Computation and Simulation 66, no. 2 (2000): 101–11. His writings appear also in classified NSA journals: Irving J. Good, “A List of Properties of Bayes-Turing Factors,” NSA Technical Journal 10, no. 2 (1965), https://www.nsa.gov/Portals/70/ documents/news-features/declassified-documents/tech-journals/list-of -properties.pdf.
24. Colin B. Burke，《并非全是魔法：20 世纪 30 年代至 60 年代自动化密码分析的早期斗争》（马里兰州米德堡：美国国家安全局密码历史中心，2002 年），第 277 页，http://archive.org/details/NSA-WasntAllMagic_2002。有关应对 20 世纪 60 年代以来不断增加的数据量的思考，请参阅 Willis Ware 的《第二计算机研究组报告，1972 年 5 月提交》，《美国国家安全局技术期刊》第 19 卷，第 1 期（1974 年）：第 21-63 页；Joseph Eachus 等人的《在 NSA 成长过程中与计算机共存（绝密的 Umbra）》，《美国国家安全局技术期刊特刊》（1972 年）：第 3-14 页。
24. Colin B. Burke, It Wasn’t All Magic: The Early Struggle to Automate Cryptanalysis, 1930s-1960s (Fort Meade, MD: Center for Cryptological History, NSA, 2002), 277, http://archive.org/details/NSA-WasntAllMagic_2002. For reflections on the need to contend with increasing volumes from the 1960s onward, see Willis Ware, “Report of the Second Computer Study Group, Submitted May 1972,” NSA Technical Journal 19, no. 1 (1974): 21–63; Joseph Eachus et al., “Growing Up with Computers at NSA (Top Secret Umbra),” NSA Technical Journal Special issue (1972): 3–14.
25. 伯克，《这并非全是魔法》，265。
25. Burke, It Wasn’t All Magic, 265.
26. Frances Allen，Paul Lasewicz 采访，2003 年 4 月 16 日，第 4 页，https://amturing.acm.org/allen_history.pdf。
26. Frances Allen, interview by Paul Lasewicz, April 16, 2003, 4, https:// amturing.acm.org/allen_history.pdf.
27. 艾伦，4-5。
27. Allen, 4–5.
28. Frances Allen，《口述历史》，Al Kossow 采访，2008 年 9 月 11 日，5。计算机历史博物馆 X5006.2009。
28. Frances Allen, Oral History, interview by Al Kossow, September 11, 2008, 5. Computer History Museum X5006.2009.
29. 伯克，《这并非全是魔法》，264。
29. Burke, It Wasn’t All Magic, 264.
30. Samuel S. Snyder，“密码组织推动的计算机进步”，计算机史年鉴2，第 1 期（1980 年）：66。
30. Samuel S. Snyder, “Computer Advances Pioneered by Cryptologic Organizations,” Annals of the History of Computing 2, no. 1 (1980): 66.
31. Samuel S. Snyder，“ABNER：ASA 计算机，第二部分：制造、操作和影响”，NSA 技术期刊，nd，83。
31. Samuel S. Snyder, “ABNER: The ASA Computer, Part II: Fabrication, Operation, and Impact,” NSA Technical Journal, n.d., 83.
32. 有关 20 世纪 70 年代和 80 年代密码学的发展（大部分仍属机密），请参阅 Thomas R. Johnson 著《冷战期间的美国密码学，1945–1989，第四册：密码学的重生 1981–1989》（美国国家安全局密码学历史中心，1999 年），第 291–292 页。
32. For developments in the 1970s and 1980s, still mostly classified, see Thomas R. Johnson, American Cryptology During the Cold War, 1945–1989, Book IV: Cryptologic Rebirth 1981–1989 (NSA Center for Cryptologic History, 1999), 291–292.
33. 摘录自《多重假设检验和贝叶斯因子（秘密）》，NSA 技术期刊16，第 3 期（1971 年）：63–80，第 71 页。
33. redacted, “Multiple Hypothesis Testing and the Bayes Factor (Secret),” NSA Technical Journal 16, no. 3 (1971): 63–80, p. 71.
34. FT Leahy，《贝叶斯因子的明显悖论（U）》，《美国国家安全局技术期刊》第 27 卷第 3 期（第 7-10 期），第 8、9 页。“对于密码学家来说，不可能存在任何先验概率分配（无论是否巧妙），从而可能对我们的计算机程序的实用性产生不利影响。”
34. F. T. Leahy, “The Apparent Paradox of Bayes Factors (U),” NSA Technical Journal 27, no. 3 (n.d.): 7–10, pp. 8, 9. “For there can exist for the cryptographer no assignment of a priori odds (whether ingenious or otherwise) that can adversely affect the usefulness of our computer program.”
35. 比较 NSATJ 中关于聚类分析的几篇论文；R51 的工作。
35. Compare several papers on cluster analysis in NSATJ; work of R51.
36. Mina Rees，《联邦计算机计划》，《科学》 112，第 2921 期（1950 年 12 月 22 日）：735；有关 Rees 在形成对数学的支持方面所起的作用的有力解释，请参阅 Alma Steingart，《公理学：数学思想和高度现代主义》（芝加哥：芝加哥大学出版社，即将出版）。
36. Mina Rees, “The Federal Computing Machine Program,” Science 112, no. 2921 (December 22, 1950): 735; for a powerful account of Rees’s role in the shaping of support for mathematics, see Alma Steingart, Axiomatics: Mathematical Thought and High Modernism (Chicago: University of Chicago Press, forthcoming).
37. Robert W. Seidel，“‘数字运算’：AEC 实验室中的计算机和物理研究”，《历史与技术》第 15 卷，第 1-2 期（1998 年 9 月 1 日）：第 54 页，https://doi.org/10.1080/07341519808581940。
37. Robert W. Seidel, “ ‘Crunching Numbers’: Computers and Physical Research in the AEC Laboratories,” History and Technology 15, no. 1–2 (September 1, 1998): 54, https://doi.org/10.1080/07341519808581940.
38. Gordon Bell、Tony Hey 和 Alex Szalay，《超越数据洪流》，《科学》 323 卷，第 5919 期（2009 年）：1297–98 页。
38. Gordon Bell, Tony Hey, and Alex Szalay, “Beyond the Data Deluge,” Science 323, no. 5919 (2009): 1297–98.
39. Vance Packard，《裸体社会》（纽约，D. McKay Co，1964 年），第 41 页，http://archive.org/details/nakedsociety00pack。
39. Vance Packard, The Naked Society (New York, D. McKay Co, 1964), 41, http://archive.org/details/nakedsociety00pack.
第七章：没有数据的智能
CHAPTER 7: INTELLIGENCE WITHOUT DATA
1.克劳德·香农写给艾琳·安格斯，1952 年 8 月 8 日，香农论文，第 1 框，引自 R. Kline 的《控制论、自动机研究和达特茅斯人工智能会议》，《IEEE 计算史年鉴》第 33 卷，第 4 期（2011 年 4 月）：8，https://doi.org/10.1109/MAHC.2010.44；有关香农在 20 世纪 50 年代初对相关努力的展望，请参阅克劳德·E·香农的《计算机和自动机》， IRE 41 会议纪要，第 10 期（1953 年 10 月）：1234–41，https://doi.org/10.1109/JRPROC .1953.274273。
1. Claude Shannon to Irene Angus, 8 Aug. 1952, Shannon Papers, box 1, quoted in R. Kline, “Cybernetics, Automata Studies, and the Dartmouth Conference on Artificial Intelligence,” IEEE Annals of the History of Computing 33, no. 4 (April 2011): 8, https://doi.org/10.1109/MAHC.2010 .44; For Shannon’s aspirational survey of efforts in the early 1950s, see Claude E. Shannon, “Computers and Automata,” Proceedings of the IRE 41, no. 10 (October 1953): 1234–41, https://doi.org/10.1109/JRPROC .1953.274273.
2.人工智能的历史在很多有影响力的通俗读物中都有所叙述。关于人工智能的更学术的历史，主要研究包括玛格丽特·A·博登的《心智作为机器：认知科学史》（纽约：牛津大学出版社，2006 年）；尼尔斯·J·尼尔森的《人工智能的探索：思想和成就的历史》（英国剑桥：剑桥大学出版社，2010 年）；罗伯托·科尔德斯基的《人工智能的发现：控制论之前和之后的行为、心智和机器》（多德雷赫特：施普林格出版社，2002 年）。
2. Much of the history of AI has been told in influential popular books. For the more academic history of AI, key major studies include Margaret A. Boden, Mind as Machine: A History of Cognitive Science (New York: Oxford University Press, 2006); Nils J. Nilsson, The Quest for Artificial Intelligence: A History of Ideas and Achievements (Cambridge, UK: Cambridge University Press, 2010); Roberto Cordeschi, The Discovery of the Artificial: Behavior, Mind and Machines Before and Beyond Cybernetics (Dordrecht: Springer, 2002).
3.请参阅 Stephanie Dick 的讨论，“数学之后：战后美国的（重新）配置思维、证明和计算”（博士论文，哈佛大学，2015 年），2-3，https://dash.harvard.edu/handle/1Z14226096。
3. See the discussion in Stephanie Dick, “After Math: (Re)Configuring Minds, Proof, and Computing in the Postwar United States” (PhD diss., Harvard University, 2015), 2–3, https://dash.harvard.edu/handle/1Z14226096.
4.艾伦·M·图灵，《计算机器和智能》，《Mind》 59，第236期（1950）：447。
4. Alan M. Turing, “Computing Machinery and Intelligence,” Mind 59, no. 236 (1950): 447.
5.图灵，《计算机器和智能》，第449页。
5. Turing, “Computing Machinery and Intelligence,” 449.
6.图灵，《计算机器和智能》，第449页。
6. Turing, “Computing Machinery and Intelligence,” 449.
7. Lucy A. Suchman，《人机重构：计划与情境行动》，第二版（剑桥和纽约：剑桥大学出版社，2007 年），第 226 页；参见 Stephanie Dick，“AfterMath：人机协作时代的证明工作”，Isis 102，第 3 期（2011 年）：495n3，https://doi.org/10.1086/661623。
7. Lucy A. Suchman, Human-Machine Reconfigurations: Plans and Situated Actions, 2nd ed. (Cambridge and New York: Cambridge University Press, 2007), 226; see Stephanie Dick, “AfterMath: The Work of Proof in the Age of Human–Machine Collaboration,” Isis 102, no. 3 (2011): 495n3, https:// doi.org/10.1086/661623.
8. John McCarthy 等人，《达特茅斯人工智能夏季研究项目提案》，1955 年 8 月 31 日，第 16 页，洛克菲勒档案中心，洛克菲勒基金会记录，项目，RG 1.2，系列 200.D，盒子 26，文件夹 219。
8. John McCarthy et al., “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence,” August 31, 1955, 16, Rockefeller Archive Center, Rockefeller Foundation records, projects, RG 1.2, series 200.D, box 26, folder 219.
9. Alma Steingart，公理化：数学思想与高度现代主义（芝加哥：芝加哥大学出版社，即将出版）。
9. Alma Steingart, Axiomatics: Mathematical Thought and High Modernism (Chicago: University of Chicago Press, forthcoming).
10. 克劳德·列维-斯特劳斯，《人的数学》，《国际社会科学通讯》第6卷第4期（1954年）：586页。讨论见Steingart的《公理体系》。
10. Claude Lévi-Strauss, “The Mathematics of Man,” International Social Science Bulletin 6, no. 4 (1954): 586. Discussed in Steingart, Axiomatics.
11. 列维-斯特劳斯，585。
11. Lévi-Strauss, 585.
12. Bruce G. Buchanan 和 Edward Hance Shortliffe，《基于规则的专家系统：斯坦福启发式编程项目的 MYCIN 实验》（马萨诸塞州雷丁：Addison-Wesley，1984 年），3，http://archive.org/details/rule basedexperts00buch。
12. Bruce G. Buchanan and Edward Hance Shortliffe, Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project (Reading, MA: Addison-Wesley, 1984), 3, http://archive.org/details/rule basedexperts00buch.
13. John McCarthy，《通用机器人只是海市蜃楼》，争议节目，BBC，1973 年 8 月 20 日，可参见《莱特希尔辩论》（1973 年）第 4 部分（共 6 部分），https://www.youtube.com/watch2v = pyU9pm1hmYs&t=266s，第 4:26 页。
13. John McCarthy, in “The General Purpose Robot is a Mirage,” Controversy Programme, BBC, August 20, 1973, available as The Light-hill Debate (1973)—Part 4 of 6, https://www.youtube.com/watch2v = pyU9pm1hmYs&t=266s, at 4:26.
14. McCarthy 等人，“达特茅斯人工智能夏季研究项目提案”。
14. McCarthy et al., “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence.”
15. 香农和麦卡锡之间的差距在 Kline 的《控制论、自动机研究和达特茅斯人工智能会议》中得到了最好的体现；Jonathan Penn 的《发明智能：20 世纪中叶美国复杂信息处理和人工智能的历史》（论文，剑桥大学，2021 年），123–24，https://doi.org/10.17863/CAM.63087。
15. The gulf between Shannon and McCarthy is best documented in Kline, “Cybernetics, Automata Studies, and the Dartmouth Conference on Artificial Intelligence”; Jonathan Penn, “Inventing Intelligence: On the History of Complex Information Processing and Artificial Intelligence in the United States in the Mid-Twentieth Century” (Thesis, University of Cambridge, 2021), 123–24, https://doi.org/10.17863/CAM.63087.
16. Penn，《发明智能》，第 134 页。在人机交互等其他计算机领域，专业心理学家发挥着更为重要的作用。参见 Sam Schirvar，《管理者的机器：秘书、心理学家和“人机交互”，1973-1983”（正在审查中）。
16. Penn, “Inventing Intelligence,” 134. In other computer fields, like Human-Computer Interaction, professional psychologists figured much more centrally. See Sam Schirvar, “Machinery for Managers: Secretaries, Psychologists, and ‘Human-Computer Interaction’, 1973–1983” (under review).
17. Pamela McCorduck，《会思考的机器：对人工智能历史与前景的个人探究》，25 周年更新版（马萨诸塞州纳蒂克：AK Peters，2004 年），114。
17. Pamela McCorduck, Machines Who Think: A Personal Inquiry into the History and Prospects of Artificial Intelligence, 25th anniversary update (Natick, MA: A.K. Peters, 2004), 114.
18. Hunter Heyck，“定义计算机：赫伯特·西蒙和官僚思维——第一部分”，IEEE 计算史年鉴30，第 2 期（2008 年 4 月）：42–51，https://doi.org/10.1109/MAHC.2008.18；Hunter Heyck，“定义计算机：赫伯特·西蒙和官僚思维——第二部分”，IEEE 计算史年鉴30，第 2 期（2008 年 4 月）：52–63，https://doi.org/10.1109/MAHC.2008.19。
18. Hunter Heyck, “Defining the Computer: Herbert Simon and the Bureaucratic Mind—Part 1,” IEEE Annals of the History of Computing 30, no. 2 (April 2008): 42–51, https://doi.org/10.1109/MAHC.2008.18; Hunter Heyck, “Defining the Computer: Herbert Simon and the Bureaucratic Mind—Part 2,” IEEE Annals of the History of Computing 30, no. 2 (April 2008): 52–63, https://doi.org/10.1109/MAHC.2008.19.
19. Allen Newell、JC Shaw 和 Herbert A. Simon，《人类问题解决理论的要素》，《心理学评论》 65，第 3 期（1958 年）：153，https://doi.org/10.1037/h0048495。
19. Allen Newell, J. C. Shaw, and Herbert A. Simon, “Elements of a Theory of Human Problem Solving,” Psychological Review 65, no. 3 (1958): 153, https://doi.org/10.1037/h0048495.
20. Penn，《发明智能》，第 45 页。
20. Penn, “Inventing Intelligence,” 45.
21. Jamie Cohen-Cole，“认知科学的反身性：科学家作为人性的模型”，《人文科学史》第 18 卷，第 4 期（2005 年 11 月 1 日）：122，https://doi.org/10.1177/0952695105058473；有关人工智能与认知科学的更广泛关系，请参阅 Cordeschi 的《人工智能的发现》。
21. Jamie Cohen-Cole, “The Reflexivity of Cognitive Science: The Scientist as Model of Human Nature,” History of the Human Sciences 18, no. 4 (November 1, 2005): 122, https://doi.org/10.1177/0952695105058473; for Al’s relationship with cognitive science more generally, Cordeschi, The Discovery of the Artificial.
22. 对于启发式方法，请比较 Ekaterina Babintseva 的《从教学计算到知识工程：Lev Landa 算法启发式理论的起源和应用》，载于《抽象与体现：计算与社会的新历史》，由 Stephanie Dick 和 Janet Abbate 编辑（马里兰州巴尔的摩：约翰霍普金斯大学出版社，2022 年）。
22. For heuristics, compare Ekaterina Babintseva, “From Pedagogical Computing to Knowledge-Engineering: The Origins and Applications of Lev Landa’s Algo-Heuristic Theory,” in Abstractions and Embodiments: New Histories of Computing and Society, ed. Stephanie Dick and Janet Abbate (Baltimore, MD: Johns Hopkins University Press, 2022).
23. 欲详细了解以数学为中心的人工智能中自动化与自由之间的矛盾，请参阅 Stephanie Dick 的《表述的政治：二十世纪美国数学中的自动化叙述》，载于《叙述科学：1800 年以来的推理、表述和认知》，由 Mary S. Morgan、Kim M. Hajek 和 Dominic J. Berry 主编（剑桥大学出版社，即将出版）。
23. For a nuanced read of tensions of automation and freedom in mathematically focused AI, see Stephanie Dick, “The Politics of Representation: Narratives of Automation in Twentieth Century American Mathematics,” in Narrative Science: Reasoning, Representing and Knowing since 1800, ed. Mary S. Morgan, Kim M. Hajek, and Dominic J. Berry (Cambridge University Press, forthcoming).
24. 关于 John McCarthy 的“具有常识的程序”的讨论，载于《思维过程的机械化；1958 年 11 月 24、25、26 和 27 日在国家物理实验室举行的研讨会论文集》，国家物理实验室（英国）（伦敦：HM 文具办公室，1961 年），第 86–87、88 页，http://archive.org/details/mechanisationoft01nati。
24. Discussion of John McCarthy, “Programs with Common Sense,” in Mechanisation of Thought Processes; Proceedings of a Symposium Held at the National Physical Laboratory on 24th, 25th, 26th and 27th November 1958, ed. National Physical Laboratory (Great Britain) (London: H.M. Stationery Office, 1961), 86–87, 88, http://archive.org/details/ mechanisationoft01nati.
25. 值得注意的是，艾莉森·亚当（Alison Adam），《人工智能：性别与思考机器》（纽约：劳特利奇，1998 年）。
25. Notably, Alison Adam, Artificial Knowing: Gender and the Thinking Machine (New York: Routledge, 1998).
26. Marvin Minsky，《迈向人工智能》，载于《计算机与思想》 ，Edward A. Feigenbaum 和 Julian Feldman 主编（纽约，McGraw-Hill，1963 年），第 428 页，http://archive.org/details/computersthought00feig。该论文最初发表于 1961 年。
26. Marvin Minsky, “Steps toward Artificial Intelligence,” in Computers and Thought, ed. Edward A. Feigenbaum and Julian Feldman (New York, McGraw-Hill, 1963), 428, http://archive.org/details/ computersthought00feig. The paper appeared originally in 1961.
27. Margaret A. Boden，“GOFAI”，《剑桥人工智能手册》，Keith Frankish 和 William M. Ramsey 编辑（英国剑桥：剑桥大学出版社，2014 年），第 89 页，https://doi.org/10.1017/CBO978 1139046855.007。
27. Margaret A. Boden, “GOFAI,” in The Cambridge Handbook of Artificial Intelligence, ed. Keith Frankish and William M. Ramsey (Cambridge, UK: Cambridge University Press, 2014), 89, https://doi.org/10.1017/CBO978 1139046855.007.
28. Jon Agar，“计算机带来了什么变化？”，《科学社会研究》第 36 卷第 6 期（2006 年 12 月 1 日）：898，https://doi.org/10.1177/0306312 706073450。
28. Jon Agar, “What Difference Did Computers Make?,” Social Studies of Science 36, no. 6 (December 1, 2006): 898, https://doi.org/10.1177/0306312 706073450.
29. Penn，《发明智能》。
29. Penn, “Inventing Intelligence.”
30. Stephanie Dick，《论模型与机器：实现有限理性》，Isis 106，第3期（2015）：630。
30. Stephanie Dick, “Of Models and Machines: Implementing Bounded Rationality,” Isis 106, no. 3 (2015): 630.
31. 詹姆斯·莱特希尔爵士，《莱特希尔报告》，《人工智能：综合调查》，1972 年 6 月，http://www.chilton-computing.org.uk/inf/literature /reports/lighthill_report/p001.htm。有关该报告的背景，请参阅乔恩·阿加尔，《科学是为了什么？莱特希尔关于人工智能的重新解释报告》，《英国科学史杂志》第 53 期第 3 期（2020 年 9 月）：289–310，https://doi.org/10.1017/S0007087420 0 00230。
31. Sir James Lighthill, “Lighthill Report,” Artificial Intelligence: A General Survey, June 1972, http://www.chilton-computing.org.uk/inf/literature /reports/lighthill_report/p001.htm. For the context of the report see Jon Agar, “What Is Science for? The Lighthill Report on Artificial Intelligence Reinterpreted,” The British Journal for the History of Science 53, no. 3 (September 2020): 289–310, https://doi.org/10.1017/S0007087420 0 00230.
32. David C. Brock，“从人工智能先前的觉醒中学习：专家系统的历史”，AI Magazine 39，第 3 期（2018 年 9 月 28 日）：3–15，https://doi.org/10.1609/aimag.v39i3.2809；David Ribes 等人，“领域的逻辑”，Social Studies of Science 49，第 3 期（2019 年 6 月 1 日）：287–91，https://doi.org/10.1177/0306312719849709；Hallam Stevens，“生物学中的商业机器——AI 在生命科学中的商业化”， IEEE Annals of the History of Computing 44，第 4 期（2019 年 6 月 1 日）：287–91，https://doi.org/10.1177/0306312719849709 01（2022 年 1 月 1 日）：8–19，https://doi.org/10.1109/MAHC.2021.3104868。
32. David C. Brock, “Learning from Artificial Intelligence’s Previous Awakenings: The History of Expert Systems,” AI Magazine 39, no. 3 (September 28, 2018): 3–15, https://doi.org/10.1609/aimag.v39i3.2809; David Ribes et al., “The Logic of Domains,” Social Studies of Science 49, no. 3 (June 1, 2019): 287–91, https://doi.org/10.1177/0306312719849709; Hallam Stevens, “The Business Machine in Biology—The Commercialization of AI in the Life Sciences,” IEEE Annals of the History of Computing 44, no. 01 (January 1, 2022): 8–19, https://doi.org/10.1109/MAHC.2021.3104868.
33. EA Feigenbaum、BG Buchanan 和 J. Lederberg，《论通用性和问题解决：使用 DENDRAL 程序的案例研究》，《机器智能》，第 6 期（1971 年）：187 页。
33. E. A. Feigenbaum, B. G. Buchanan, and J. Lederberg, “On Generality and Problem Solving: A Case Study Using the DENDRAL Program,” Machine Intelligence, no. 6 (1971): 187.
34. Marvin Minsky 和 Seymour Papert，《人工智能进展报告》，1971 年，人工智能备忘录 AIM-252，https://web.media.mit.edu/~minsky/papers/PR1971.html。
34. Marvin Minsky and Seymour Papert, “Progress Report on Artificial Intelligence,” 1971, Artificial Intelligence Memo AIM-252, https://web.media .mit. edu/~ minsky/p ap ers/PR 1971 .html.
35. Ira Goldstein 和 Seymour Papert，“人工智能、语言和知识研究”，认知科学1，第 1 期（1977 年 1 月 1 日）：85，https://doi.org/10.1016/S03 64–0213 (77)8000 6–2。
35. Ira Goldstein and Seymour Papert, “Artificial Intelligence, Language, and the Study of Knowledge,” Cognitive Science 1, no. 1 (January 1, 1977): 85, https://doi.org/10.1016/S03 64–0213 (77)8000 6–2.
36. Joseph Adam November，《生物医学计算：美国生命的数字化》（巴尔的摩：约翰霍普金斯大学出版社，2012 年），259–68。
36. Joseph Adam November, Biomedical Computing: Digitizing Life in the United States (Baltimore: Johns Hopkins University Press, 2012), 259–68.
37.Buchanan 和 Shortliffe，基于规则的专家系统，16。
37. Buchanan and Shortliffe, Rule-Based Expert Systems, 16.
38. 有关瓶颈，请参阅 Stephanie A. Dick，“Coded Conduct: Making MACSYMA Users and the Automation of Mathematics”，BJHS Themes 5（2020 年编辑）：205–24，https://doi.org/10.1017/bjt.2020.10；Edward Feigenbaum，Oral History，Nils Nilsson 访谈，2007 年 20、27、62–63，http://archive.computerhistory.org/resources/access/text/2013/05/102702002–05-01-acc.pdf；DE Forsythe，“Engineering Knowledge: The Construction of Artificial Intelligence”，Social Studies of Science 23，no. 3 (1993 年 8 月 1 日)：445–77，https://doi.org/10.1177/0306312793023003002。
38. For the bottleneck, see Stephanie A. Dick, “Coded Conduct: Making MACSYMA Users and the Automation of Mathematics,” BJHS Themes 5 (ed. 2020): 205–24, https://doi.org/10.1017/bjt.2020.10; Edward Feigenbaum, Oral History, interview by Nils Nilsson, 20, 27 2007, 62–63, http:// archive.computerhistory.org/resources/access/text/2013/05/102702002–05 -01-acc.pdf; D. E. Forsythe, “Engineering Knowledge: The Construction of Knowledge in Artificial Intelligence,” Social Studies of Science 23, no. 3 (August 1, 1993): 445–77, https://doi.org/10.1177/0306312793023003002.
39. JR Quinlan，“通过从大量实例中归纳发现规则”，《微电子时代的专家系统》，Donald Michie 主编（爱丁堡：爱丁堡大学出版社，1979 年），第 168 页。
39. J. R. Quinlan, “Discovering Rules by Induction from Large Collections of Examples,” in Expert Systems in the Micro-Electronic Age, ed. Donald Michie (Edinburgh: Edinburgh University Press, 1979), 168.
40.Donald Michie，《专家系统访谈》，《专家系统》 2 卷，第 1 期（1985 年）：第 22 页。
40. Donald Michie, “Expert Systems Interview,” Expert Systems 2, no. 1 (1985): 22.
41. 有关专家系统的隐藏成功，请参阅 Stevens，“生物学中的商业机器——人工智能在生命科学中的商业化”。
41. For the hidden successes of expert systems, see Stevens, “The Business Machine in Biology—The Commercialization of AI in the Life Sciences.”
42. Jacob T. Schwartz，《人工智能的极限》（纽约：纽约大学 Courant 数学科学研究所，1986 年），30，http://archive.org/details/limitsofartifici00schw。
42. Jacob T. Schwartz, The Limits of Artificial Intelligence (New York: Courant Institute of Mathematical Sciences, New York University, 1986), 30, http:// archive.org/details/limitsofartifici00schw.
43. Y. Bar-Hillel 对英国国家物理实验室麦卡锡的《思维过程的机械化》的评论；24、25、26 和 27 日在国家物理实验室举行的研讨会论文集 1958 年 11 月（伦敦，HM Stationery off.，1961 年），85，http://archive.org/details/mechanisationoftOlnati。
43. Comments by Y. Bar-Hillel on McCarthy, National Physical Laboratory (Great Britain), Mechanisation of Thought Processes; Proceedings of a Symposium Held at the National Physical Laboratory on 24th, 25th, 26th and 27th November 1958 (London, H.M. Stationery off., 1961), 85, http://archive.org /details/mechanisationoftOlnati.
44. 有关纽曼，请参阅 B. Jack Copeland，《马克斯·纽曼——数学家、密码破译者、计算机先驱》，载《巨人：布莱切利园密码破译计算机的秘密》，B. Jack Copeland 主编（英国牛津：牛津大学出版社，2006 年），第 176–88 页。
44. For Newman, see B. Jack Copeland, “Max Newman—Mathematician, Code Breaker, Computer Pioneer,” in Colossus: The Secrets of Bletchley Park’s Codebreaking Computers, ed. B. Jack Copeland (Oxford, UK: Oxford University Press, 2006), 176–88.
45. EA Newman，《非数学数据处理分析》，《思维过程的机械化；1958 年 11 月 24、25、26 和 27 日在国家物理实验室举行的研讨会论文集》，英国国家物理实验室 (伦敦：HM 文具办公室，1961 年)，第 866 页，http://archive.org/details/mechanisationoft02nati。
45. E. A. Newman, “An Analysis of Non Mathematical Data Processing,” in Mechanisation of Thought Processes; Proceedings of a Symposium Held at the National Physical Laboratory on 24th, 25th, 26th and 27th November 1958, ed. National Physical Laboratory (Great Britain) (London: H.M. Stationery Office, 1961), 866, http://archive.org/details/mechanisationoft02nati.
46. 纽曼，875。
46. Newman, 875.
47. Richard O. Duda 和 Peter E Hart，《模式分类和场景分析》（纽约：Wiley，1973 年）。
47. Richard O. Duda and Peter E Hart, Pattern Classification and Scene Analysis (New York: Wiley, 1973).
第 8 章：数量、种类和速度
CHAPTER 8: VOLUME, VARIETY, AND VELOCITY
1. R. Blair Smith，《Robina Mapstone 的口述历史》（查尔斯·巴贝奇研究所，1980 年 5 月），27、29，http://conservancy.umn.edu/handle/11299/107637。
1. R. Blair Smith, Oral History by Robina Mapstone (Charles Babbage Institute, May 1980), 27, 29, http://conservancy.umn.edu/handle/11299/107637.
2. Martin Campbell-Kelly，《从航空预订到刺猬索尼克：软件行业历史》（马萨诸塞州剑桥：麻省理工学院出版社，2003 年），第 43 页，http://www.loc.gov/catdir/toc/fy035/2002075351.html。另请参阅 https:// www.ibm.com/ibm/history/ibm100/us/en/icons/sabre/team/
2. Martin Campbell-Kelly, From Airline Reservations to Sonic the Hedgehog: A History of the Software Industry (Cambridge, MA: MIT Press, 2003), 43, http://www.loc.gov/catdir/toc/fy035/2002075351.html. See also https:// www.ibm.com/ibm/history/ibm100/us/en/icons/sabre/team/
3. RW Parker，《SABRE 系统》，Datamation 11（1965 年 9 月）：49。参见 Campbell-Kelly，《从航空预订到刺猬索尼克》，第 41-45 页。
3. R. W. Parker, “The SABRE System,” Datamation 11 (September 1965): 49. See Campbell-Kelly, From Airline Reservations to Sonic the Hedgehog, 41–45.
4.隐私保护研究委员会，《信息社会中的个人隐私：隐私保护研究委员会报告》。（华盛顿：该委员会：由文件主管出售，美国政府印刷局，1977年），第4页。
4. Privacy Protection Study Commission, Personal Privacy in an Information Society: The Report of the Privacy Protection Study Commission. (Washington: The Commission: For sale by the Supt. of Docs., US Govt. Print. Off., 1977), 4.
5.引自 Thomas J. Misa，《数字国家：明尼苏达州计算产业的故事》（明尼阿波利斯：明尼苏达大学出版社，2013 年），第 64 页。
5. Quotation from Thomas J. Misa, Digital State: The Story of Minnesota’s Computing Industry (Minneapolis: University of Minnesota Press, 2013), 64.
6. James W. Cortada，《数字洪流：信息技术在美国、欧洲和亚洲的传播》（纽约：牛津大学出版社，2012 年），第 49 页。
6. James W. Cortada, The Digital Flood: The Diffusion of Information Technology across the U.S., Europe, and Asia (New York: Oxford University Press, 2012), 49.
7. Samuel S. Snyder，《密码组织推动的计算机进步》，《计算机历史年鉴》第 2 卷，第 1 期（1980 年）：第 60–70 页，第 65 页。“SOLO 是美国第一台完全晶体管化的计算机。”
7. Samuel S. Snyder, “Computer Advances Pioneered by Cryptologic Organizations,” Annals of the History of Computing 2, no. 1 (1980): 60–70, at 65. “SOLO holds the distinction of being the first completely transistorized computer in the United States.”
8. Eckert-Mauchly 计算机公司 (EMCC)，《UNIVAC 系统广告》，1948 年，第 2、5 页，https://www.computerhistory.org/revolution/early-computer-companies/5/103/447?position = 0。请参阅 Arthur L. Norberg 的《计算机与商业：Eckert-Mauchly 计算机公司、工程研究协会和雷明顿兰德公司 1946-1957 年的技术和管理研究》（马萨诸塞州剑桥：麻省理工学院出版社，2005 年），第 185-186 页中的讨论。
8. Eckert-Mauchly Computer Corporation (EMCC), “UNIVAC System Advertisement,” 1948, 2, 5, https://www.computerhistory.org/revolution/early -computer-companies/5/103/447?position = 0. See the discussion in Arthur L. Norberg, Computers and Commerce: A Study of Technology and Management at Eckert-Mauchly Computer Company, Engineering Research Associates, and Remington Rand, 1946–1957 (Cambridge, MA: MIT Press, 2005), 185–86.
9. Norberg，《计算机与商业》，第 191 页。有关磁带不可靠性的更多信息，请参阅 Thomas Haigh 的《镀铬制表机：1954-1958 年电子革命的制度化》，《IEEE 计算史年鉴》第 23 卷第 4 期（2001 年）：第 86、88 页。
9. Norberg, Computers and Commerce, 191. For more on the unreliability of tape, see Thomas Haigh, “The Chromium-Plated Tabulator: Institutionalizing an Electronic Revolution, 1954–1958,” IEEE Annals of the History of Computing 23, no. 4 (2001): 86, 88.
10. J. Abbate，《重新编码性别：女性在计算领域不断变化的参与度》（马萨诸塞州剑桥：麻省理工学院出版社，2012 年），第 37-38 页。
10. J. Abbate, Recoding Gender: Women’s Changing Participation in Computing (Cambridge, MA: MIT Press, 2012), 37–38.
11. 有关这些努力，请参阅 James W. Cortada 的“1945–1995 年美国公司的数字计算机”，IEEE 计算史年鉴18，第 2 期（1996 年）：18–29；Haigh，“镀铬制表机”。
11. For these efforts, see James W. Cortada, “Commercial Applications of the Digital Computer in American Corporations, 1945–1995,” IEEE Annals of the History of Computing 18, no. 2 (1996): 18–29; Haigh, “The Chromium-Plated Tabulator.”
12. JM Juran，引自 Richard G. Canning，《商业和工业电子数据处理》（纽约：Wiley，1956 年），第 316 页，https://catalog.hathitrust.org/Record/001118357。
12. J. M. Juran, quoted in Richard G. Canning, Electronic Data Processing for Business and Industry (New York: Wiley, 1956), 316, https://catalog .hathitrust.org/Record/001118357.
13. 施乐广告，Datamation，11（1965 年 9 月），第 76 页。
13. Xerox advertisement, Datamation, 11 (September 1965), p. 76.
14. Control Data Corporation 广告，Datamation，11（1965 年 9 月），第 87 页。
14. Control Data Corporation advertisement, Datamation, 11 (September 1965), p. 87.
15. 引自 Haigh 著《镀铬制表机》，第 97 页。
15. Quoted in Haigh, “The Chromium-Plated Tabulator,” 97.
16. Paul Edwards，《巨大的机器：计算机模型、气候数据和全球变暖的政治》（马萨诸塞州剑桥：麻省理工学院出版社，2010 年），第 111 页。
16. Paul Edwards, A Vast Machine: Computer Models, Climate Data, and the Politics of Global Warming (Cambridge, MA: MIT Press, 2010), 111.
17. Martha Poon，“记分卡作为消费者信贷工具：Fair, Isaac & Company Incorporated 案例”，《社会学评论》 55，第 2 期增刊（2007 年 10 月）：284–306，https://doi.org/10.1111/j.1467–954X.2007.00740.x。
17. Martha Poon, “Scorecards as Devices for Consumer Credit: The Case of Fair, Isaac & Company Incorporated,” The Sociological Review 55, no. 2_ suppl (October 2007): 284–306, https://doi.org/10.1111/j.1467–954X.2007 .00740.x.
18. Josh Lauer，《消费者编码：信用报告和信用评分的计算机化》，载于《Creditworthy》（纽约：哥伦比亚大学出版社，2017 年），第 183 页，https://doi.org/10.7312/laue16808–009。有关信用评估的更长期历史，请参阅 Rowena Olegario 的《信用文化：在美国商业中嵌入信任和透明度》（马萨诸塞州剑桥：哈佛大学出版社，2006 年）。
18. Josh Lauer, “Encoding the Consumer: The Computerization of Credit Reporting and Credit Scoring,” in Creditworthy (New York: Columbia University Press, 2017), 183, https://doi.org/10.7312/laue16808–009. For the longer history of gauging credit, see Rowena Olegario, A Culture of Credit : Embedding Trust and Transparency in American Business (Cambridge, MA: Harvard University Press, 2006).
19. 有关信用卡的发展和衡量信用度的形式，请参阅 Louis Hyman 的《债务人国家：美国赤字史》（新泽西州普林斯顿：普林斯顿大学出版社，2011 年），第 7 章。
19. For the developments of credit cards and forms of gauging creditworthiness, see Louis Hyman, Debtor Nation: The History of America in Red Ink (Princeton, NJ: Princeton University Press, 2011), ch 7.
20. “Datamation：编辑读物：老大哥”，Datamation 11（1965年10月）：第23页。
20. “Datamation: Editor’s Readout: Big Brother,” Datamation 11 (October 1965): 23.
21. Packard，《赤裸的社会》，第 41 页。我们在隐私方面的工作深受 Sarah Elizabeth Igo 的《已知公民：现代美国的隐私史》（马萨诸塞州剑桥：哈佛大学出版社，2018 年）。Igo 教授的评论极大地丰富了本章。
21. Packard, The Naked Society, 41. Our work on privacy is deeply indebted to Sarah Elizabeth Igo, The Known Citizen: A History of Privacy in Modern America (Cambridge, MA: Harvard University Press, 2018). Comments from Professor Igo sharpened this chapter immensely.
22. 帕卡德，《赤裸社会》，41页。
22. Packard, The Naked Society, 41.
23. Stanton Wheeler 编，《记录：美国生活中的文件和档案》（新泽西州新不伦瑞克：Transaction Books，1976 年），第 19-20 页。
23. Stanton Wheeler, ed., On Record: Files and Dossiers in American Life (New Brunswick, NJ: Transaction Books, 1976), 19–20.
24. Arthur R. Miller，《对隐私的侵犯：计算机、数据库和档案》（纽约：新美国图书馆，1972 年），第 22 页。
24. Arthur R. Miller, The Assault on Privacy: Computers, Data Banks, and Dossiers (New York: New American Library, 1972), 22.
25. 美国国会参议院政府运作委员会广告隐私和信息特设小组委员会，隐私：个人数据的收集、使用和计算机化：美国参议院政府运作委员会隐私和信息系统特设小组委员会与司法委员会宪法权利小组委员会联合听证会，第九十三届国会，第二届会议...1974 年 6 月 18、19 和 20 日（华盛顿：美国政府印刷厂，1974 年），I：53。
25. United States Congress Senate Committee on Government Operations Ad Hoc Subcommittee on Privacy and Information, Privacy: The Collection, Use, and Computerization of Personal Data: Joint Hearings Before the Ad Hoc Subcommittee on Privacy and Information Systems of the Committee on Government Operations and the Subcommittee on Constitutional Rights of the Committee on the Judiciary, United States Senate, Ninety-Third Congress, Second Session... June 18, 19, and 20, 1974 (Washington: US Govt. Print. Off, 1974), I:53.
26. Dan Bouk，“国家数据中心和数据替身的崛起”，自然科学历史研究48，第 5 期（2018 年 11 月 1 日）：627–36，https://doi.org/10.1525/hsns.2018.48.5.627；Igo，《已知公民》；Priscilla M. Regan，《隐私立法：技术、社会价值观和公共政策》（教堂山：北卡罗来纳大学出版社，1995 年），71–73。
26. Dan Bouk, “The National Data Center and the Rise of the Data Double,” Historical Studies in the Natural Sciences 48, no. 5 (November 1, 2018): 627–36, https://doi.org/10.1525/hsns.2018.48.5.627; Igo, The Known Citizen; Priscilla M. Regan, Legislating Privacy: Technology, Social Values, and Public Policy (Chapel Hill: University of North Carolina Press, 1995), 71–73.
27. 美国国会参议院政府运作委员会隐私和信息特设小组委员会，参议院隐私特设委员会，II：1739。
27. United States Congress Senate Committee on Government Operations Ad Hoc Subcommittee on Privacy and Information, Senate Ad Hoc Committee Privacy, II: 1739.
28. 美国国会参议院政府运作委员会隐私和信息特设小组委员会，II：1741。
28. United States Congress Senate Committee on Government Operations Ad Hoc Subcommittee on Privacy and Information, II: 1741.
29. 美国国会参议院政府运作委员会隐私和信息特设小组委员会艾伦·威斯汀的证词，I：77-78。
29. Testimony of Alan Westin, United States Congress Senate Committee on Government Operations Ad Hoc Subcommittee on Privacy and Information, I:77–78.
30. 引用自美国国会参议院政府运作委员会隐私和信息特设小组委员会，参议院隐私特设委员会，I：651。
30. Quoted in United States Congress Senate Committee on Government Operations Ad Hoc Subcommittee on Privacy and Information, Senate Ad Hoc Committee Privacy, I:651.
31. 美国国家银行美国国会参议院政府运作委员会隐私和信息特设小组委员会声明，参议院隐私特设委员会。I：606。
31. Statement of the National Bank Americard, United States Congress Senate Committee on Government Operations Ad Hoc Subcommittee on Privacy and Information, Senate Ad Hoc Committee Privacy. I:606.
32. 美国国会参议院政府运作委员会隐私和信息特设小组委员会，参议院隐私特设委员会，I：658。
32. United States Congress Senate Committee on Government Operations Ad Hoc Subcommittee on Privacy and Information, Senate Ad Hoc Committee Privacy, I:658.
33. Igo，《已知公民》，362–63。
33. Igo, The Known Citizen, 362–63.
34. 司法委员会法院、公民自由和司法行政小组委员会，1984 年，《公民自由和国家安全状态：司法委员会法院、公民自由和司法行政小组委员会听证会》。众议院第九十八届国会，第一、二届会议，1984 年：《公民自由和国家安全状态》，1983 年 11 月 2、3 日和 1984 年 1 月 24 日、4 月 5 日和 9 月 26 日。（华盛顿特区：美国邮政总局，1986 年），第 267-79 页。
34. Subcommittee on Courts, Civil Liberties, and the Administration of Justice of the Committee on the Judiciary, 1984, Civil Liberties and the National Security State: Hearings before the Subcommittee on Courts, Civil Liberties, and the Administration of Justice of the Committee on the Judiciary. House of Representatives, Ninety-Eighth Congress, First and Second Sessions, on 1984: Civil Liberties and the National Security State, November 2, 3, 1983 and January 24, April 5, and September 26, 1984. (Washington, DC: US G.P.O., 1986), 267–79.
35. 司法委员会法院、公民自由和司法行政小组委员会，1984 年，《公民自由和国家安全》，第 294-95 页。
35. Subcommittee on Courts, Civil Liberties, and the Administration of Justice of the Committee on the Judiciary, 1984, Civil Liberties and the National Security State, 294–95.
36.技术评估、电子记录系统和个人隐私 办公室，11。
36. Office of Technology Assessment, Electronic Record Systems and Individual Privacy, 11.
37.技术评估、电子记录系统和个人隐私 办公室，39，40。
37. Office of Technology Assessment, Electronic Record Systems and Individual Privacy, 39, 40.
38. 美国隐私保护研究委员会，《信息社会中的个人隐私》，533（原文全部为斜体）。
38. United States and Privacy Protection Study Commission, Personal Privacy in an Information Society, 533 (original all in italics).
39. 例如，请参阅 Eubanks 的《自动化不平等》。
39. See, e.g., Eubanks, Automating Inequality.
40. Bobbie Johnson 和 Las Vegas，《Facebook 创始人称隐私不再是社会规范》，《卫报》 ，2010 年 1 月 11 日，sec. Technology，https://www.theguardian.com/technology/2010/jan/11/facebook-privacy。
40. Bobbie Johnson and Las Vegas, “Privacy No Longer a Social Norm, Says Facebook Founder,” The Guardian, January 11, 2010, sec. Technology, https://www.theguardian.com/technology/2010/jan/11/facebook-privacy.
41. W. Lee Burge，“信息的自由流动：信用经济的关键”，载于美国国会参议院政府运作委员会隐私和信息特设小组委员会，《隐私：个人数据的收集、使用和计算机化：政府运作委员会隐私和信息系统特设小组委员会与司法委员会宪法权利小组委员会联合听证会》，美国参议院，第九十三届国会，第二届会议……1974 年 6 月 18、19 和 20 日（华盛顿特区：美国政府印刷厂，1974 年），I：650。
41. W. Lee Burge, “The Free Flow of Information: Key to Our Credit Economy” in United States Congress Senate Committee on Government Operations Ad Hoc Subcommittee on Privacy and Information, Privacy: The Collection, Use, and Computerization of Personal Data: Joint Hearings Before the Ad Hoc Subcommittee on Privacy and Information Systems of the Committee on Government Operations and the Subcommittee on Constitutional Rights of the Committee on the Judiciary, United States Senate, Ninety-Third Congress, Second Session...June 18, 19, and 20, 1974 (Washington, DC: US Govt. Print. Off, 1974), I:650.
42. Jennifer Barrett Glasgow，“Acxiom，致众议员 Edward J. Markey 的信”，2012 年 8 月 15 日，http://markey.house.gov/sites/markey.house.gov/files/documents/Acxiom.pdf。
42. Jennifer Barrett Glasgow, “Acxiom, Letter to Representative Edward J. Markey,” August 15, 2012, http://markey.house.gov/sites/markey.house .gov/files/documents/Acxiom.pdf.
43. New 和 Castro，“政策制定者如何促进算法问责制”，2。
43. New and Castro, “How Policymakers Can Foster Algorithmic Accountability,” 2.
44. Paul Baran，《立法、隐私和EDUCOM》，《EDUCOM：大学间通信委员会公报》，1968年12月3日，第3页。
44. Paul Baran, “Legislation, Privacy and EDUCOM,” EDUCOM: Bulletin of the Interuniversity Communications Council, December 3, 1968, 3.
45. Paul Baran，“论未来的计算机时代：美国人性格和工程师角色的改变，或在数字热潮中稍加谨慎”，兰德公司论文（兰德公司，1968 年），第 14 页，https://www.rand.org/pubs/p ap ers/P3780.html。
45. Paul Baran, “On the Future Computer Era: Modification of the American Character and the Role of the Engineer, or, A Little Caution in the Haste to Number,” RAND Paper (RAND Corporation, 1968), 14, https://www.rand . org/pubs/p ap ers/P3 78 0.html.
46. 罗伯特·诺齐克，《无政府、国家和乌托邦》（纽约：Basic Books，1974），32-33。
46. Robert Nozick, Anarchy, State, and Utopia (New York: Basic Books, 1974), 32–33.
47. Daniel T. Rodgers，《断裂时代》（马萨诸塞州剑桥：哈佛大学出版社，2011 年），190。
47. Daniel T. Rodgers, Age of Fracture (Cambridge, MA: Harvard University Press, 2011), 190.
48. 米尔顿·弗里德曼，“企业的社会责任是增加利润（1970 年）”，载于《企业伦理与公司治理》，Walther Ch Zimmerli、Markus Holzinger 和 Klaus Richter 编辑（柏林、海德堡：Springer，2007 年），第 178 页，https://doi.org/10.1007/978–3–540–70818–6_14。
48. Milton Friedman, “The Social Responsibility of Business Is to Increase Its Profits (1970),” in Corporate Ethics and Corporate Governance, ed. Walther Ch Zimmerli, Markus Holzinger, and Klaus Richter (Berlin, Heidelberg: Springer, 2007), 178, https://doi.org/10.1007/978–3–540–70818–6_14.
49. 参见 Jodi L. Short，“监管改革中的偏执风格”，《黑斯廷斯法律杂志》第 63 期（2012 年）：第 633–94 页；Julie E. Cohen，“真相与权力之间：信息资本主义的法律建构”（纽约：牛津大学出版社，2019 年），第 189 页；Amy Kapczynski，“信息资本主义的法律”，《耶鲁法律杂志》 ，2020 年，第 1491 页。
49. See Jodi L. Short, “The Paranoid Style in Regulatory Reform,” Hastings Law Journal 63 (2012): 633–94; Julie E. Cohen, Between Truth and Power: The Legal Constructions of Informational Capitalism (New York: Oxford University Press, 2019), 189; Amy Kapczynski, “The Law of Informational Capitalism,” The Yale Law Journal, 2020, 1491.
50. Priscilla M. Regan，《隐私立法：技术、社会价值和公共政策》（教堂山：北卡罗来纳大学出版社，1995 年），第 4 页。
50. Priscilla M. Regan, Legislating Privacy: Technology, Social Values, and Public Policy (Chapel Hill: University of North Carolina Press, 1995), 4.
51. Oscar H. Gandy，《全景分类：个人信息的政治经济学》，《传播与文化产业批判研究》（Boulder，CO：Westview，1993 年）。
51. Oscar H. Gandy, The Panoptic Sort: A Political Economy of Personal Information, Critical Studies in Communication and in the Cultural Industries (Boulder, CO: Westview, 1993).
52. Matthew Crain，《利润高于隐私：监控广告如何征服互联网》（明尼阿波利斯：明尼苏达大学出版社，2021 年），第 20 页。
52. Matthew Crain, Profit Over Privacy: How Surveillance Advertising Conquered the Internet (Minneapolis: University of Minnesota Press, 2021), 20.
53. Meg Leta Jones，“Cookies：争议的遗产”，《互联网历史》 4，第 1 期（2020 年 1 月 2 日）：87–104，https://doi.org/10.1080/24701475.2020.1725852。
53. Meg Leta Jones, “Cookies: A Legacy of Controversy,” Internet Histories 4, no. 1 (January 2, 2020): 87–104, https://doi.org/10.1080/24701475.2020 .1725852.
54. 引自 Joshua Quittner 所著《快乐的恶作剧者去华盛顿》 ， 《连线》，2021 年 5 月 14 日访问，https://www.wired.com/1994/06/eff/；讨论见 Fred Turner 所著《从反主流文化到赛博文化：斯图尔特·布兰德、全球网络和数字乌托邦主义的兴起》（芝加哥：芝加哥大学出版社，2006 年），第 219 页。
54. Quoted in Joshua Quittner, “The Merry Pranksters Go to Washington,” Wired, accessed May 14, 2021, https://www.wired.com/1994/06/eff/; discussed in Fred Turner, From Counterculture to Cyberculture: Stewart Brand, the Whole Earth Network, and the Rise of Digital Utopianism (Chicago: University of Chicago Press, 2006), 219.
55. Turner，《从反主流文化到赛博文化》，261。
55. Turner, From Counterculture to Cyberculture, 261.
56. 保罗·萨宾，《公民：对大政府的攻击和美国自由主义的重塑》（纽约：诺顿，2021 年）。
56. Paul Sabin, Public Citizens: The Attack on Big Government and the Remaking of American Liberalism (New York: Norton, 2021).
57. 参见 Kapczynski，《信息资本主义法则》，1493–99 年；还引用了 Cohen 的《真相与权力之间》。
57. See Kapczynski, “The Law of Informational Capitalism,” 1493–99; drawing on Cohen, Between Truth and Power.
58. 比较从广义的职业责任理解向狭义的民权关注的重大转变，Megan Finn 和 Quinn DuPont 在“从封闭的世界话语到数字乌托邦主义：计算机专业人员社会责任组织负责任计算的变迁（1981-1992）”中讨论了这一转变，互联网历史4，第 1 期（2020 年 1 月 2 日）：6-31，https://doi.org/10.1080/24701475.2020.1725851。
58. Compare the dramatic shift from a broad understanding of professional responsibility toward a narrower civil rights focus, discussed in Megan Finn and Quinn DuPont, “From Closed World Discourse to Digital Utopianism: The Changing Face of Responsible Computing at Computer Professionals for Social Responsibility (1981–1992),” Internet Histories 4, no. 1 (January 2, 2020): 6–31, https://doi.org/10.1080/24701475.2020.1725851.
59. 引自秘密决定，删除了姓名和日期，第 63 页，引自《修订备忘录意见》第 8-9 页。
59. Quotation from secret decision with redacted name and date, p. 63, quoted in Amended Memorandum Opinion at 8–9.
60. 修订后的备忘录意见第8页。
60. Amended Memorandum Opinion at 8.
61. Felten，“Edward W. Felten 教授在美国公民自由联盟等人诉James R. Clapper 等人案中的声明”，第 8 页。
61. Felten, “Declaration of Professor Edward W. Felten in ACLU et al. v. James R. Clapper et al.,” 8.
62. 匿名，《经验教训。对[删除]的采访》，第 1 页。
62. Anonymous, “Lessons Learned. Interview with [Redacted],” 1.
63. 有关同意的挑战，例如，请参阅 Frank Pasquale，《许可作为数据治理》，哥伦比亚大学 Knight 第一修正案研究所，2021 年 9 月 28 日，https://knightcolumbia.org/content/licensure-as-data-governance。
63. For challenges for consent, see, for example, Frank Pasquale, “Licensure as Data Governance,” Knight First Amendment Institute at Columbia University, September 28, 2021, https://knightcolumbia.org/content/licensure-as -data-governance.
64. R Allen Wilkinson 等人，第一次人口普查光学字符识别系统会议，NIST IR 4912（马里兰州盖瑟斯堡：国家标准与技术研究所，1992 年），1，https://doi.org/10.6028/NIST.IR .4912。
64. R Allen Wilkinson et al., The First Census Optical Character Recognition System Conference, NIST IR 4912 (Gaithersburg, MD: National Institute of Standards and Technology, 1992), 1, https://doi.org/10.6028/NIST.IR .4912.
65.Wilkinson 等人，4。
65. Wilkinson et al., 4.
66. Wilkinson 等人，2.
66. Wilkinson et al., 2.
第九章：机器、学习
CHAPTER 9: MACHINES, LEARNING
1. Pat Langley，“机器学习科学的变化”，机器学习82，第 3 期（2011 年 3 月）：277，https://doi.org/10.1007/s10994–011–5242-y。有关预测和机器学习的历史，请参阅 Adrian Mackenzie，“预测的产生：机器学习想要什么？”欧洲文化研究杂志18，第 4-5 期（2015 年）：429–45；Ann Johnson，“预测的理性和经验文化”，数学作为一种工具，Johannes Lenhard 和 Martin Carrier 编辑，第 327 卷（瑞士 Cham：Springer International Publishing，2017 年），23–35，https://doi.org/10.1007/978–3–319–54469–4_2； Adrian Mackenzie，《机器学习者：数据实践考古学》（马萨诸塞州剑桥：麻省理工学院出版社，2018 年）；Aaron Plasek，“论真正撰写机器学习史的残酷性”，IEEE 计算史年鉴38，第 4 期（2016 年 12 月）：6–8，https://doi.org/10.1109/MAHC.2016.43；Aaron Mendon-Plasek，“机械化意义和机器学习：为什么教机器评判世界成为可以想象和更可取的做法”，《机器学习的文化生活：对批判性人工智能研究的入侵》，Jonathan Roberge 和 Michael Castelle 编辑（瑞士卡姆：Springer International Publishing，2021 年），第 31–78 页，https://doi.org/10.1007/978–3–030–56286–1_2； Cosma Rohilla Shalizi，《统计学习范式和机器学习领域的形成，约 1985-2000 年》（2020 年），手稿正在准备中，可向作者索取。
1. Pat Langley, “The Changing Science of Machine Learning,” Machine Learning 82, no. 3 (March 2011): 277, https://doi.org/10.1007/s10994–011–5242-y. For histories of prediction and machine learning, see Adrian Mackenzie, “The Production of Prediction: What Does Machine Learning Want?,” European Journal of Cultural Studies 18, no. 4–5 (2015): 429–45; Ann Johnson, “Rational and Empirical Cultures of Prediction,” in Mathematics as a Tool, ed. Johannes Lenhard and Martin Carrier, vol. 327 (Cham, Switzerland: Springer International Publishing, 2017), 23–35, https://doi.org/10 .1007/978–3–319–54469–4_2; Adrian Mackenzie, Machine Learners: Archaeology of a Data Practice (Cambridge, MA: MIT Press, 2018); Aaron Plasek, “On the Cruelty of Really Writing a History of Machine Learning,” IEEE Annals of the History of Computing 38, no. 4 (December 2016): 6–8, https:// doi.org/10.1109/MAHC.2016.43; Aaron Mendon-Plasek, “Mechanized Significance and Machine Learning: Why It Became Thinkable and Preferable to Teach Machines to Judge the World,” in The Cultural Life of Machine Learning: An Incursion into Critical AI Studies, ed. Jonathan Roberge and Michael Castelle (Cham, Switzerland: Springer International Publishing, 2021), 31–78, https://doi.org/10.1007/978–3–030–56286–1_2; Cosma Rohilla Shalizi, “The Formation of the Statistical Learning Paradigm and the Field of Machine Learning, c. 1985–2000” (2020), manuscript in preparation, available on request from the author.
2. P. Langley 和 JG Carbonell，《机器学习方法》，《美国信息科学学会杂志》第 35 卷，第 5 期（1984 年 9 月 1 日）：第 306-16 页，第 306 页。
2. P. Langley and J. G. Carbonell, “Approaches to Machine Learning,”Journal of the American Society for Information Science 35, no. 5 (September 1, 1984): 306–16, at 306.
3. Rosenblatt，“感知器：一种感知和识别自动机（PARA 项目）”，1，https://blogs.umass.edu/brain-wars/files/2016/03 /rosenblatt-1957.pdf。
3. Rosenblatt, “The Perceptron: A Perceiving and Recognizing Automaton (Project PARA),” 1, https://blogs.umass.edu/brain-wars/files/2016/03 /rosenblatt-1957.pdf.
4. Jonathan Penn，《发明智能：20 世纪中叶美国复杂信息处理和人工智能的历史》（论文，剑桥大学，2021 年），第 96–98 页，https://doi.org/10.17863/CAM.63087。有关 Rosenblatt 和同期经济项目的信息，请查看 Orit Halpern，《未来将无法计算：神经网络、新自由主义和反动政治》，《批判性探究》第 48 卷，第 2 期（2022 年 1 月 1 日）：第 334–59 页，https://doi .org/10.1086/717313。
4. Jonathan Penn, “Inventing Intelligence: On the History of Complex Information Processing and Artificial Intelligence in the United States in the Mid-Twentieth Century” (Thesis, University of Cambridge, 2021), 96–98, https://doi.org/10.17863/CAM.63087. For Rosenblatt and contemporaneous economic projects, check out Orit Halpern, “The Future Will Not Be Calculated: Neural Nets, Neoliberalism, and Reactionary Politics,” Critical Inquiry 48, no. 2 (January 1, 2022): 334–59, https://doi .org/10.1086/717313.
5. “海军新设备边做边学；心理学家展示旨在阅读和变得更聪明的计算机胚胎”，《纽约时报》，1958 年 7 月 8 日，http://trmesmachrne.nytrmes.com/trmesmachrne/1958/07/08/834r7341 .html。
5. “New Navy Device Learns by Doing; Psychologist Shows Embryo of Computer Designed to Read and Grow Wiser,” New York Times, July 8, 1958, http://trmesmachrne.nytrmes.com/trmesmachrne/1958/07/08/834r7341 .html.
6. Herbert A. Simon，“机器为什么要学习？” ，《机器学习：一种人工智能方法》，Ryszard S. Michalski、Jaime G. Carbonell 和 Tom M. Mitchell 编辑，《符号计算》（柏林、海德堡：Springer，1983 年），第 32 页，https://doi.org/10.1007/978–3–662–12405–5_2。
6. Herbert A. Simon, “Why Should Machines Learn?,” in Machine Learning: An Artificial Intelligence Approach, ed. Ryszard S. Michalski, Jaime G. Carbonell, and Tom M. Mitchell, Symbolic Computation (Berlin, Heidelberg: Springer, 1983), 32, https://doi.org/10.1007/978–3–662–12405–5_2.
7.有关这些努力的概述，特别是围绕斯坦福的活动，请参阅 Nils J. Nilsson 的《人工智能的探索：思想和成就的历史》（英国剑桥：剑桥大学出版社，2010 年），第 4 章。
7. For an overview of these efforts, with special attention to activities around Stanford, see Nils J. Nilsson, The Quest for Artificial Intelligence: A History of Ideas and Achievements (Cambridge, UK: Cambridge University Press, 2010), ch. 4.
8.有关政府与学术界之间这一研究“灰色地带”的重要性，请参阅 Joy Rohde 著《以专业知识武装自己：冷战期间美国社会研究的军事化》（纽约州伊萨卡：康奈尔大学出版社，2013 年）。
8. For the importance of this “gray area” of research between government and academia, see Joy Rohde, Armed with Expertise: The Militarization of American Social Research during the Cold War (Ithaca, NY: Cornell University Press, 2013).
9. Laveen N. Kanal， 《模式识别》中的“前言” ，编辑。 Laveen N. Kanal（华盛顿特区：Thompson Book Co，1968 年），xi。
9. Laveen N. Kanal, “Preface,” in Pattern Recognition, ed. Laveen N. Kanal (Washington, DC: Thompson Book Co, 1968), xi.
10. G. Nagy，“模式识别的最新进展”，IEEE 56 论文集，第 5 期（1968 年 5 月）：836-63，https://doi.org/10.n09/PROC.1968.6414。
10. G. Nagy, “State of the Art in Pattern Recognition,” Proceedings of the IEEE 56, no. 5 (May 1968): 836–63, https://doi.org/10.n09/PROC.1968.6414.
11. Xiaochang Li，《“没有比更多数据更重要的数据”：自动语音识别和算法文化的形成》，摘自James Evans 和 Adrian Johns 编著的《Osiris，超越工艺和代码》，即将出版，抓住了这一关键点。
11. Xiaochang Li, “ ‘There’s No Data Like More Data’: Automatic Speech Recognition and the Making of Algorithmic Culture,” in Osiris, “Beyond Craft and Code,” ed. James Evans and Adrian Johns, forthcoming, captures this crucial point.
12. Mendon-Plasek，《机械化意义与机器学习》，第 2-3 页。Michael Castelle 正在完成损失函数的主要历史研究。
12. Mendon-Plasek, “Mechanized Significance and Machine Learning,” 2–3. Michael Castelle is finishing a major history of the loss function.
13. J. McCarthy 等人，《达特茅斯人工智能夏季研究项目提案》，1955 年 8 月 31 日，洛克菲勒档案中心，洛克菲勒基金会记录，项目，RG 1.2，系列 200.D，盒子 26，文件夹 219。
13. J. McCarthy et al., “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence,” August 31, 1955, Rockefeller Archive Center, Rockefeller Foundation records, projects, RG 1.2, series 200.D, box 26, folder 219.
14. James Lighthill 爵士，《Lighthill 报告》，《人工智能：综合调查》，1972 年 6 月，第 3 节，http://www.chilton-computing.org.uk/inf/literature/rep orts/lighthill_rep ort/p001.htm。
14. Sir James Lighthill, “Lighthill Report,” Artificial Intelligence: A General Survey, June 1972, §3, http://www.chilton-computing.org.uk/inf/literature /rep orts/lighthill_rep ort/p 0 01.htm.
15. Rodney A. Brooks，《没有代表的智能》，《人工智能》 47 卷，1-3 期（1991 年）：140。
15. Rodney A. Brooks, “Intelligence without Representation,” Artificial Intelligence 47, no. 1–3 (1991): 140.
16. 有关专家系统的隐藏成功，请参阅 Hallam Stevens，《生物学中的商业机器——生命科学中人工智能的商业化》，《IEEE 计算史年鉴》第 44 卷，第 01 期（2022 年 1 月 1 日）：8–19，https://doi.org/10.n09/MAHC.2021.3104868。
16. For the hidden success of expert systems, see Hallam Stevens, “The Business Machine in Biology—The Commercialization of AI in the Life Sciences,” IEEE Annals of the History of Computing 44, no. 01 (January 1, 2022): 8–19, https://doi.org/10.n09/MAHC.2021.3104868.
17. Keki B. Irani 等人，“将机器学习应用于半导体制造”，IEEE Expert 8，no. 1（1993）：41。
17. Keki B. Irani et al., “Applying Machine Learning to Semiconductor Manufacturing,” IEEE Expert 8, no. 1 (1993): 41.
18. Alain Desrosières，《Prouver et gouverner：uneanalyze politique des statistiques publiques》（巴黎：Découverte，2014 年），第 1 章。 9.
18. Alain Desrosières, Prouver et gouverner: une analyse politique des statistiques publiques (Paris: Découverte, 2014), ch. 9.
19. Hayashi Chikio和M. Takahashi，“Kagakusi to Kagakusha：Hayashi Chikiosi Kōkai Intabyu”，KōdōKeiryōgaku 31，no。 2（2004）：107-24；引用并翻译于 Joonwoo Son，“日本的数据科学”（未出版的 MS，哥伦比亚大学，社会学，2016 年 5 月）。
19. Hayashi Chikio and M. Takahashi, “Kagakusi to Kagakusha: Hayashi Chikiosi Kōkai Intabyu,” KōdōKeiryōgaku 31, no. 2 (2004): 107–24; quoted and translated in Joonwoo Son, “Data Science in Japan” (Unpublished MS, Columbia University, Sociology, May 2016).
20. Vladimir Naumovich Vapnik，《基于经验数据的依赖性估计》（1982）；《经验推断科学：2006 年后记》第 2 版（纽约：Springer，2006 年），第 415 页。
20. Vladimir Naumovich Vapnik, Estimation of Dependences Based on Empirical Data (1982); Empirical Inference Science: Afterword of 2006, 2nd ed. (New York: Springer, 2006), 415.
21. 请参阅“ИПУ PAH 研究所历史”，2017 年 7 月 7 日访问，http://www.ipu.ru/en/node/12744。
21. See “History of the Institute | ИПУ PAH,” accessed July 7, 2017, http:// www.ipu.ru/en/node/12744.
22. Vapnik 经常严厉批评大多数其他数据分析的数学和理论不足。如需了解，请参阅 Léon Bottou 的“In Hindsight: Doklady Akademii Nauk SSSR, 181 (4), 1968”，载于《Empirical Inference》（Springer，2013 年），第 3-5 页，http://link.springer.com/chapter /10.1007/978–3–642–4113 6–6_1。
22. Vapnik has often been highly critical of the mathematical and theoretical insufficiencies of most other data analysis. For an appreciation, see Léon Bottou, “In Hindsight: Doklady Akademii Nauk SSSR, 181 (4), 1968,” in Empirical Inference (Springer, 2013), 3–5, http://link.springer.com/chapter /10.1007/978–3–642–4113 6–6_1.
23. Xiaochang Li，“占卜引擎：文本预测的媒体历史”（纽约大学，2017 年）；Xiaochang Li 和 Mara Mills，“声音特征：从语音识别到机器语音识别”，《科技与文化》第 60 卷，第 2 期（2019 年 6 月 18 日）：S129–60，https://doi.org/10.1353/tech.2019.0066。
23. Xiaochang Li, “Divination Engines: A Media History of Text Prediction” (NYU, 2017); Xiaochang Li and Mara Mills, “Vocal Features: From Voice Identification to Speech Recognition by Machine,” Technology and Culture 60, no. 2 (June 18, 2019): S129–60, https://doi.org/10.1353/tech.2019.0066.
24. 有关引人入胜的自传体叙述，请参阅 Terrence J. Sejnowski，《深度学习革命》（马萨诸塞州剑桥：麻省理工学院出版社，2018 年）； Yann LeCun，Quand la machine apprend：La Révolution des Neurones artificiels et de l'apprentissage profond（巴黎：Odile Jacob，2019）；有关精彩的新闻报道，请参阅 John Markoff，《Machines of Loving Grace：The Quest for Common Ground Between Humans and Robots》，2016 年；以及 Dominique Cardon、Jean-Philippe Cointet 和 Antoine Mazières 的出色论述，“神经元的复兴：机器归纳的发明和人工智能的争议”，Réseaux n° 211，第 1 期。 5（2018）：173，https://doi.org/10.3917/res.211.0173。
24. For compelling autobiographical accounts, see Terrence J. Sejnowski, The Deep Learning Revolution (Cambridge, MA: MIT Press, 2018); Yann LeCun, Quand la machine apprend: La Révolution des neurones artificiels et de l’apprentissage profond (Paris: Odile Jacob, 2019); for a fine journalistic account, see John Markoff, Machines of Loving Grace: The Quest for Common Ground between Humans and Robots, 2016; and the excellent account Dominique Cardon, Jean-Philippe Cointet, and Antoine Mazières, “La revanche des neurones: L’invention des machines inductives et la controverse de l’intelligence artificielle,” Réseaux n° 211, no. 5 (2018): 173, https://doi.org/10.3917/res.211.0173.
25. 虽然该方法的不同版本早已发表，但关键的变革性研究包括 David E. Rumelhart 和 James L. McClelland 的《通过误差传播学习内部表征》，载于《并行分布式处理：认知微观结构的探索：基础》（马萨诸塞州剑桥：麻省理工学院出版社，1987 年），第 318–62 页，http://ieeexplore.ieee.org/document/6302929；David E. Rumelhart、Geoffrey E. Hinton 和 Ronald J. Williams 的《通过反向传播误差学习表征》，《自然》第 323 卷，第 6088 期（1986 年 10 月）：第 533–36 页，https://doi.org/10.1038/323533a0； Y. LeCun 等人，“反向传播应用于手写邮政编码识别”，《神经计算》1，第 4 期（1989 年 12 月）：541-51，https://doi.org/10.1162/neco.1989.1.4.541；PJ Werbos，“穿越时间的反向传播：它的作用和方法”，《IEEE 论文集》 78，第 10 期（1990 年 10 月）：1550-60，https://doi.org/10.1109/5.58337。
25. While versions of the approach were published earlier, the key transformative studies include David E. Rumelhart and James L. McClelland, “Learning Internal Representations by Error Propagation,” in Parallel Distributed Processing: Explorations in the Microstructure of Cognition: Foundations (Cambridge, MA: MIT Press, 1987), 318–62, http://ieeexplore .ieee.org/document/6302929; David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams, “Learning Representations by Back-Propagating Errors,” Nature 323, no. 6088 (October 1986): 533–36, https://doi.org/10 .1038/323533a0; Y. LeCun et al., “Backpropagation Applied to Handwritten Zip Code Recognition,” Neural Computation 1, no. 4 (December 1989): 541–51, https://doi.org/10.1162/neco.1989.1.4.541; P. J. Werbos, “Backpropagation through Time: What It Does and How to Do It,” Proceedings of the IEEE 78, no. 10 (October 1990): 1550–60, https://doi.org/10.1109/5.58337.
26. 有关这一时期的许多令人难忘的故事，请参阅 LeCun 的《当机器发挥作用时》。
26. For many memorable stories of this period, see LeCun, Quand la machine apprend.
27. 来自 Cardon、Cointet 和 Mazières 的匿名受访者，“神经元的复仇”，21。
27. From an anonymous interviewee in Cardon, Cointet, and Mazières, “La revanche des neurones,” 21.
28. Leo Breiman 和 Nong Shang，“重生之树”，nd，https://www.stat.berkeley.edu/~breiman/BAtrees.pdf。
28. Leo Breiman and Nong Shang, “Born Again Trees,” n.d., https://www.stat .berkeley.edu/~breiman/BAtrees.pdf.
29. 有关神经网络的复兴，请参阅 Yann LeCun、Yoshua Bengio 和 Geoffrey Hinton 的《深度学习》，《自然》 521 卷，第 7553 期（2015 年 5 月 27 日）：43644，https://doi.org/10.1038/nature14539。
29. For the revival of neural nets, see Yann LeCun, Yoshua Bengio, and Geoffrey Hinton, “Deep Learning,” Nature 521, no. 7553 (May 27, 2015): 43644, https://doi.org/10.1038/nature14539.
30. Alex Krizhevsky、Ilya Sutskever 和 Geoffrey E. Hinton，“使用深度卷积神经网络进行 ImageNet 分类”，载于《神经信息处理系统进展》，第 25 卷（Curran Associates, Inc.，2012 年），https://papers.nips.cc/paper/2012/hash/c399862d3b9d6 b76c8436e924a68c45b-Abstract.html。
30. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” in Advances in Neural Information Processing Systems, vol. 25 (Curran Associates, Inc., 2012), https://papers.nips.cc/paper/2012/hash/c399862d3b9d6 b76c8436e924a68c45b-Abstract.html.
31. Olga Russakovsky 等人，“ImageNet 大规模视觉识别挑战赛”， 国际计算机视觉杂志115，第 3 期（2015 年 12 月 1 日）：211–52，https://doi.org/10.1007/s11263–015–0816-y。
31. Olga Russakovsky et al., “ImageNet Large Scale Visual Recognition Challenge,” International Journal of Computer Vision 115, no. 3 (December 1, 2015) : 211–52, https://doi.org/10.1007/s11263–015–0816-y.
32. Fei-Fei Li，“众包、基准测试和其他有趣的事情”，https://web.archive.org/web/20121110041643/http://www.image-net.org/papers/ImageNet_2010.pdf；Hao Su、Jia Deng 和 Li Fei-Fei，“用于视觉对象检测的众包注释”，nd，第 7 页。
32. Fei-Fei Li, “Crowdsourcing, Benchmarking & Other Cool Things,” https:// web. archive. org/web/20121110041643/http://www.image-net.org/papers/ ImageNet_2010.pdf; Hao Su, Jia Deng, and Li Fei-Fei, “Crowdsourcing Annotations for Visual Object Detection,” n.d., 7.
33. 有关该数据集的深层问题和争议，请参阅 Kate Crawford 和 Trevor Paglen，“Excavating AI”，2019 年 9 月 19 日，https://excavating.ai。
33. For the deep problems and controversies over this data set, see Kate Crawford and Trevor Paglen, “Excavating AI,” September 19, 2019, https:// excavating.ai.
34. Cardon、Cointet 和 Mazières，“神经元的复仇”。
34. Cardon, Cointet, and Mazières, “La revanche des neurones.”
35. 环境损失的确切范围存在很大争议。研究这种大规模计算的环境和基础设施成本的学者包括 Mél Hogan，《数据流和水资源困境：犹他州数据中心》，《大数据与社会》 2，第 2 期（2015 年 12 月 1 日）：2053951715592429，https://doi.org/10.1177/2053951715592429；Nathan Ensmenger，《计算的环境史》，《技术与文化》 59，第 4 期（2018 年）：S7–33，https://doi.org/10.1353/tech.2018.0148； Kate Crawford，《人工智能地图集：权力、政治和人工智能的地球成本》（康涅狄格州纽黑文：耶鲁大学出版社，2021 年），https://doi.org/10.12987/9780300252392，第 1 章。
35. The precise scope of the environmental toll is much debated. Scholars looking at the environmental and infrastructural costs of this massive computing include Mél Hogan, “Data Flows and Water Woes: The Utah Data Center,” Big Data & Society 2, no. 2 (December 1, 2015): 2053951715592429, https://doi.org/10.1177/2053951715592429; Nathan Ensmenger, “The Environmental History of Computing,” Technology and Culture 59, no. 4 (2018): S7–33, https://doi.org/10.1353/tech.2018.0148; Kate Crawford, Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence (New Haven, CT: Yale University Press, 2021), https://doi .org/10.12987/9780300252392, ch. 1.
36. Meredith Whittaker，“捕获的高昂成本”，《互动》第 28 卷，第 6 期（2021 年 11 月）：52，https://doi.org/10.1145/3488666。
36. Meredith Whittaker, “The Steep Cost of Capture,” Interactions 28, no. 6 (November 2021): 52, https://doi.org/10.1145/3488666.
37. MI Jordan 和 TM Mitchell，“机器学习：趋势、观点和前景”，Science 349，第 6245 期（2015 年 7 月 17 日）：255–60，https://doi.org/10.1126/science.aaa8415。
37. M. I. Jordan and T. M. Mitchell, “Machine Learning: Trends, Perspectives, and Prospects,” Science 349, no. 6245 (July 17, 2015): 255–60, https://doi.org/10.112 6/science.aaa8415.
38. 乔丹和米切尔。
38. Jordan and Mitchell.
39. Langley，“机器学习的科学变化”，第 278 页。
39. Langley, “The Changing Science of Machine Learning,” 278.
40. 请参阅 Jenna Burrell，“机器如何‘思考’：理解机器学习算法中的不透明度”，《大数据与社会》第 3 卷，第 1 期（2016 年 1 月 5 日）：4–5，https://doi.org/10.1177/2053951715622512。
40. See Jenna Burrell, “How the Machine ‘Thinks’: Understanding Opacity in Machine Learning Algorithms,” Big Data & Society 3, no. 1 (January 5, 2016) : 4–5, https://doi.org/10.1177/2053951715622512.
41. “Netflix 奖：评论规则”，2007 年 2 月 2 日，https://web.archive.org/web /20070202023620/http://www.netflixprize.com:80/rules。
41. “Netflix Prize: Review Rules,” February 2, 2007, https://web.archive.org/web /20070202023620/http://www.netflixprize.com:80/rules.
42. 引自 Steve Lohr 的文章“Netflix 的 100 万美元研究交易，或许是其他公司的典范”，《纽约时报》，2009 年 9 月 22 日，sec. Technology，https://www.nytimes.com/2009/09/22/technology/internet /22netflix.html。
42. Quoted in Steve Lohr, “A $1 Million Research Bargain for Netflix, and Maybe a Model for Others,” New York Times, September 22, 2009, sec. Technology, https://www.nytimes.com/2009/09/22/technology/internet /22netflix.html.
43. David Donoho，“50 年数据科学”，《计算与图形统计杂志》第 26 卷，第 4 期（2017 年 10 月 2 日）：752，https://doi.org/10.1080 /10618600.2017.1384734。
43. David Donoho, “50 Years of Data Science,” Journal of Computational and Graphical Statistics 26, no. 4 (October 2, 2017): 752, https://doi.org/10.1080 /10618600.2017.1384734.
44.Donoho ，752–53。
44. Donoho, 752–53.
45. 引自 Lohr 的文章“Netflix 的 100 万美元研究交易，或许可以成为其他公司的典范。”
45. Quoted in Lohr, “A $1 Million Research Bargain for Netflix, and Maybe a Model for Others.”
第 10 章：数据科学
CHAPTER 10: THE SCIENCE OF DATA
1.艾伦·金斯伯格，《嚎叫》，text/html，诗歌基金会（诗歌基金会，2021 年 8 月 12 日），https://www.poetryfoundation.org/，https://www .poetryfoundation.org/poems/49303/howl。
1. Allen Ginsberg, “Howl,” text/html, Poetry Foundation (Poetry Foundation, August 12, 2021), https://www.poetryfoundation.org/, https://www .poetryfoundation.org/poems/49303/howl.
2. Ashlee Vance，《我们信任广告》，《彭博商业周刊》 ，第 4521 期（2017 年 5 月 8 日）：第 6 页。
2. Ashlee Vance, “In Ads We Trust,” Bloomberg Businessweek, no. 4521 (May 8, 2017): 6.
3. Chris Anderson，《长尾理论》，《连线》 ，2004 年 10 月，https://www.wired.com/2004/10/tail/。
3. Chris Anderson, “The Long Tail,” Wired, October 2004, https://www .wired. c om/2004/10/tail/.
4. Gregory Zuckerman，《解决市场问题的人：吉姆·西蒙斯如何发起量化革命》（纽约：企鹅出版社，2019 年）。关于语音识别技术的发展与金融市场之间的深层关系，Xiaochang Li，《占卜引擎：文本预测的媒体历史》（博士论文，纽约大学，2017 年）。
4. Gregory Zuckerman, The Man Who Solved the Market: How Jim Simons Launched the Quant Revolution (New York: Penguin, 2019). For deep relationship between the development of speech recognition technologies and financial markets, Xiaochang Li, “Divination Engines: A Media History of Text Prediction” (PhD Thesis, NYU, 2017).
5. Ognjenka Goga Vukmirovic 和 Shirley M. Tilghman，《探索基因组空间》，《自然》 405，第 6788 期（2000 年 6 月）：820，https://doi.org/10.1038/35015690。另请参阅 Hallam Stevens，《生命无序：数据驱动的生物信息学史》（芝加哥：芝加哥大学出版社，2013 年）；Sabina Leonelli，《以数据为中心的生物学：一项哲学研究》（芝加哥：芝加哥大学出版社，2016 年），第 18 页。
5. Ognjenka Goga Vukmirovic and Shirley M. Tilghman, “Exploring Genome Space,” Nature 405, no. 6788 (June 2000): 820, https://doi.org/10 .1038/35015690. Generally see Hallam Stevens, Life Out of Sequence: A Data-Driven History of Bioinformatics (Chicago: University of Chicago Press, 2013); Sabina Leonelli, Data-Centric Biology: A Philosophical Study (Chicago: University of Chicago Press, 2016), 18.
6. Cathy O'Neil，“数据科学：工具与工艺”，Mathbabe（博客），2011 年 10 月 4 日，https://mathbabe.org/2011/10/04/data-science-tools-vs-craft/。
6. Cathy O’Neil, “Data Science: Tools vs. Craft,” Mathbabe (blog), October 4, 2011, https://mathbabe.org/2011/10/04/data-science-tools-vs-craft/.
7. Cosma Rohilla Shalizi，“新‘数据科学家’不过是旧‘统计学家’的翻版”，2011 年 12 月 4 日，https://web.archive.org/web/20111204161344/ http://cscs.umich.edu/~crshalizi/weblog/805.html。
7. Cosma Rohilla Shalizi, “New ‘Data Scientist’ Is But Old ‘Statistician’ Writ Large,” December 4, 2011, https://web.archive.org/web/20111204161344/ http://cscs.umich.edu/~crshalizi/weblog/805.html.
8.所罗门·库尔巴克（Solomon Kullback）致图基（Tukey），1959 年 3 月 13 日，美国哲学学会 [APS] 图基论文，Ms 117，系列 I：美国：NSA。
8. Solomon Kullback to Tukey, 13.3.1959, American Philosophical Society [APS] Tukey Papers, Ms 117, Series I: US: NSA.
9.霍华德·巴洛 (Howard Barlow) 代表所罗门·库尔巴克 (Solomon Kullback) 致约翰·图基 (John Tukey)，1959 年 6 月 4 日，APS Tukey 论文，Ms 117，系列 I：美国：NSA。
9. Howard Barlow for Solomon Kullback to John Tukey, 6.4.1959, APS Tukey Papers, Ms 117, Series I: US: NSA.
10. 所罗门·库尔巴克（Solomon Kullback）致图基（Tukey），1959 年 3 月 13 日，APS 图基论文，Ms 117，系列 I：美国：NSA。
10. Solomon Kullback to Tukey, 13.3.1959, APS Tukey Papers, Ms 117, Series I: US: NSA.
11. John W. Tukey，《数据分析的未来》，《数理统计年鉴》第 33 卷，第 1 期（1962 年）：6，斜体部分为原文所在。
11. John W. Tukey, “The Future of Data Analysis,” The Annals of Mathematical Statistics 33, no. 1 (1962): 6, italics his.
12. Luisa T. Fernholz 等人，“与 John W. Tukey 和 Elizabeth Tukey 的对话”，《统计科学》第 15 卷，2000 年，第 80-81 页。
12. Luisa T. Fernholz et al., “A Conversation with John W. Tukey and Elizabeth Tukey,” Statistical Science 15, 2000, 80–81.
13. John W. Tukey，《探索性数据分析》（Addison Wesley，1977 年），第 2-3 页；有关 Tukey 的项目，请参阅 Alexander Campolo，“视觉引领：数据、可视化和信息世界观的诞生”（博士论文，纽约大学，2019 年），第 186-188 页。
13. John W. Tukey, Exploratory Data Analysis (Addison Wesley, 1977), 2–3; for Tukey’s project see Alexander Campolo, “Steering by Sight: Data, Visualization, and the Birth of an Informational Worldview” (PhD diss., New York University, 2019), 186–88.
14.Tukey ，探索性数据分析，56。
14. Tukey, Exploratory Data Analysis, 56.
15. John M. Chambers，“更多还是更少的统计数据：未来研究的选择”，《统计与计算》 3，第 4 期（1993 年）：182。
15. John M. Chambers, “Greater or Lesser Statistics: A Choice for Future Research,” Statistics and Computing 3, no. 4 (1993): 182.
16. 钱伯斯，182。
16. Chambers, 182.
17. 钱伯斯，183。
17. Chambers, 183.
18. 本世纪第一个十年末，有一些具有影响力的文章认为数据的充裕度将带来理解科学的新方式，请参见 Chris Anderson 的《理论的终结：数据洪流使科学方法过时》 (The End of Theory: The Data Durge Makes the Scientific Method Obsolete)，《连线》 (Wired)，2008 年，http://archive.wired.com/science/discoveries/magazine/16-07/pb_theory；Tony Hey、Stewart Tansley 和 Kristin Tolle 的《第四范式：数据密集型科学发现》，《第四范式：数据密集型科学发现》 (Microsoft Research，2009 年)，https://www.microsoft.com/en-us/research/publication/fourth-paradigm-data-intensive-scientific-discovery/。
18. For examples of influential articles near the end of the first decade of this century arguing that the abundance of data would usher in new ways of understanding science, see Chris Anderson, “The End of Theory: The Data Deluge Makes the Scientific Method Obsolete,” Wired, 2008, http:// archive.wired.com/science/discoveries/magazine/16–07/pb_theory; Tony Hey, Stewart Tansley, and Kristin Tolle, The Fourth Paradigm: Data-Intensive Scientific Discovery, The Fourth Paradigm: Data-Intensive Scientific Discovery (Microsoft Research, 2009), https://www.microsoft.com/ en-us/research/publication/fourth-paradigm-data-intensive-scientific -discover y/.
19. “约翰·M·钱伯斯”，https://awards.acm.org/award_winners/chambers_6640862。
19. “John M. Chambers,” https://awards.acm.org/award_winners/ chambers_6640862.
20. William S. Cleveland，“数据科学：扩大统计领域的技术领域”，《国际统计评论》第 69 卷，第 1 期（2001 年 4 月）：23，https://doi.org/10.2307/1403527。
20. William S. Cleveland, “Data Science: An Action Plan for Expanding the Technical Areas of the Field of Statistics,” International Statistical Review / Revue Internationale de Statistique 69, no. 1 (April 2001): 23, https://doi .org/10.2307/1403527.
21. 该领域的核心会议是国际超大规模数据库会议，于 1975 年首次召开。
21. The central conference in the field, first convened in 1975, was the International Conference on Very Large Data Bases.
22. Usama Fayyad，《挖掘数据库：面向知识发现算法》，《数据工程技术委员会公报》 21 卷，第 1 期（1998 年）：48。
22. Usama Fayyad, “Mining Databases: Towards Algorithms for Knowledge Discovery,” Bulletin of the Technical Committee on Data Engineering 21, no. 1 (1998): 48.
23. 请参阅 Usama M. Fayyad、Gregory Piatetsky-Shapiro 和 Padhraic Smyth，《从数据挖掘到知识发现：概述》，《知识发现和数据挖掘的发展》 （加州门洛帕克：AAAI/MIT 出版社，1996 年），第 1-34 页。
23. See Usama M. Fayyad, Gregory Piatetsky-Shapiro, and Padhraic Smyth, “From Data Mining to Knowledge Discovery: An Overview,” in Advances in Knowledge Discovery and Data Mining (Menlo Park, CA: AAAI/MIT Press, 1996), 1–34.
24. Matthew L. Jones，“查询档案：从 Apriori 到 Page-Rank 的数据库挖掘”，《档案中的科学：过去、现在、未来》，Lorraine Daston 主编（芝加哥：芝加哥大学出版社，2016 年），第 311–28 页。
24. Matthew L. Jones, “Querying the Archive: Database Mining from Apriori to Page-Rank,” in Science in the Archives: Pasts, Presents, Futures, ed. Lorraine Daston (Chicago: University of Chicago Press, 2016), 311–28.
25. Shawn Thelen、Sandra Mottner 和 Barry Berman，“数据挖掘：通往营销金矿之路”，《商业视野》第 47 期（2004 年）：26，https://doi . org/10.1016/j.bushor.2004.09.005。
25. Shawn Thelen, Sandra Mottner, and Barry Berman, “Data Mining: On the Trail to Marketing Gold,” Business Horizons 47 (2004): 26, https://doi . org/10.1016/j.bushor.2004.09.005.
26. Patrick O. Brown 和 David Botstein，“利用 DNA 微阵列探索基因组新世界”，Nature Genetics 21（1999 年 1 月 1 日）：26，https://doi.org/10.1038/4462。
26. Patrick O. Brown and David Botstein, “Exploring the New World of the Genome with DNA Microarrays,” Nature Genetics 21 (January 1, 1999): 26, https://doi.org/10.1038/4462.
27. 截至 1998 年底的研讨会名单可在 http://web.archive.org/web/1999 0116232602/http://www.almaden.ibm.com/cs/quest/seminars.html 和 http://web.archive.org/web/19980210042739/http://www.almaden.ibm.com/cs/quest/seminars-hist.html 上查阅。
27. The list of seminars as of the end of 1998 are available at http://web.archive.org/web/1999 0116232602/http://www.almaden.ibm.com/cs/quest/seminars.html and http://web.archive.org/web/19980210042739/http://www.almaden.ibm.com/cs/quest/seminars-hist.html.
28. MIDAS 的网页保存在 http://infolab.stanford.edu/midas/；可以在 Yahoo e-groups 上找到数据挖掘组的列表服务。请参阅 Jeffrey D. Ullman 的“斯坦福的 MIDAS 数据挖掘项目”，载于《数据库工程与应用》，1999 年。IDEAS'99。国际研讨会论文集，1999 年，第 460-64 页，http://ieeexplore.ieee.org/xpls/abs_ all.jsp?arnumber = 787298。
28. The webpage for MIDAS is preserved at http://infolab.stanford.edu/midas/; a listserv of the data mining group can be found on Yahoo e-groups. See Jeffrey D. Ullman, “The MIDAS Data-Mining Project at Stanford,” in Database Engineering and Applications, 1999. IDEAS’99. International Symposium Proceedings, 1999, 460–64, http://ieeexplore.ieee.org/xpls/abs_ all.jsp?arnumber = 787298.
29. 该材料的印刷版本出现在 Sergey Brin、Rajeev Mot-wani 和 Terry Winograd 的《你可以用口袋里的网络做什么》中，数据工程通讯21（1998 年）：37-47。
29. A printed version of this material appeared as Sergey Brin, Rajeev Mot-wani, and Terry Winograd, “What Can You Do with a Web in Your Pocket,” Data Engineering Bulletin 21 (1998): 37–47.
30. http://infolab.stanford.edu/midas/。
30. http://infolab.stanford.edu/midas/.
31. Thomas Haigh，《网络缺失的环节：搜索引擎和门户》，《互联网与美国商业》 ，William Aspray 和 Paul Ceruzzi 编辑（马萨诸塞州剑桥：麻省理工学院出版社，2008 年），第 160-61 页。Sergey Brin 和 Lawrence Page，《大型超文本网络搜索的剖析》Engine”，第七届国际万维网会议（WWW 1998），1998 年，http://ilpubs.stanford.edu:8090/361/，§3.1。“...信息检索系统的大部分研究都是针对小型且控制良好的同质集合，例如科学论文集或相关主题的新闻报道。”
31. Thomas Haigh, “The Web’s Missing Links: Search Engines and Portals,” in The Internet and American Business, ed. William Aspray and Paul Ceruzzi (Cambridge, MA: MIT Press, 2008), 160–61. Sergey Brin and Lawrence Page, “The Anatomy of a Large-Scale Hypertextual Web Search Engine,” in Seventh International World-Wide Web Conference (WWW 1998) , 1998, http://ilpubs.stanford.edu:8090/361/, §3.1. “. . . most of the research on information retrieval systems is on small well controlled homogeneous collections such as collections of scientific papers or news stories on a related topic.”
32. Sergey Brin 和 Lawrence Page，《动态数据挖掘：通过抽样探索大型规则空间》，技术报告（Stanford InfoLab，1999 年 11 月），2，http://ilpubs.stanford.edu:8090/424/。
32. Sergey Brin and Lawrence Page, “Dynamic Data Mining: Exploring Large Rule Spaces by Sampling,” Technical Report (Stanford InfoLab, November 1999) , 2, http://ilpubs.stanford.edu:8090/424/.
33. Brin 和 Page，“大型超文本网络搜索引擎的剖析”，§4.2。
33. Brin and Page, “The Anatomy of a Large-Scale Hypertextual Web Search Engine,” §4.2.
34. Brin 和 Page，§4.2。
34. Brin and Page, §4.2.
35. 海量数据流的统计分析(美国国家科学院国家研究委员会，2004)，8-9。
35. Statistical Analysis of Massive Data Streams (National Research Council of the National Academies, 2004), 8–9.
36. Alon Halevy、Fernando Pereira 和 Peter Norvig，“数据的不合理有效性”，《智能系统》，IEEE 24，第 2 期（2009 年 4 月）：8-12，https://doi.org/10.1109/MIS.2009.36。
36. Alon Halevy, Fernando Pereira, and Peter Norvig, “The Unreasonable Effectiveness of Data,” Intelligent Systems, IEEE 24, no. 2 (April 2009): 8–12, https://doi.org/10.1109/MIS.2009.36.
37. 删节版，“面对情报的未来（U）对美国国家安全局副局长 William P. Crowell 的采访（U）”，Cryptolog 22，第 2 期（1996 年）：1-5。
37. Redacted, “Confronting the Intelligence Future (U) An Interview with William P. Crowell, NSA’s Deputy Director (U),” Cryptolog 22, no. 2 (1996): 1–5.
38. Paul Burkhardt 和 Chris Waring，“NSA 大图实验”。
38. Paul Burkhardt and Chris Waring, “An NSA Big Graph Experiment.”
39. 国家安全局 职位编号：1034503
39. NSA Job ID: 1034503
40. 删节版，“20 世纪 80 年代至 21 世纪的 NSA 文化——SID 视角”，《密码学季刊》第 30 卷第 4 期（2011 年）：第 84 页。项目符号以连续的散文形式呈现。
40. Redacted, “NSA Culture, 1980s to the 21st Century—a SID Perspective,” Cryptological Quarterly 30, no. 4 (2011): 84. Bulleted points are rendered as continuous prose.
41. Catherine D'Ignazio 和 Lauren F. Klein，《数据女权主义》（马萨诸塞州剑桥：麻省理工学院出版社，2020 年），173。
41. Catherine D’Ignazio and Lauren F. Klein, Data Feminism (Cambridge, MA: MIT Press, 2020), 173.
42. Antonio A. Casilli，《En attendant les robots: enquête sur le travail du clic》（巴黎：Éditions du Seuil，2019 年），14。
42. Antonio A. Casilli, En attendant les robots: enquête sur le travail du clic (Paris: Éditions du Seuil, 2019), 14.
43. Sarah T. Roberts，《屏幕背后：社交媒体阴影下的内容审核》（康涅狄格州纽黑文：耶鲁大学出版社，2019 年）。
43. Sarah T. Roberts, Behind the Screen: Content Moderation in the Shadows of Social Media (New Haven, CT: Yale University Press, 2019).
44. Mary L. Gray 和 Siddharth Suri，《幽灵工作：如何阻止硅谷建立新的全球下层阶级》（波士顿：霍顿·米夫林哈考特，2019 年）。
44. Mary L. Gray and Siddharth Suri, Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass (Boston: Houghton Mifflin Harcourt, 2019).
45. 卡西利，《机器人服务员》，17。
45. Casilli, En attendant les robots, 17.
46. Lilly Irani，《微工作的文化工作》，《新媒体与社会》第 17 卷，第 5 期（2015 年 5 月）：723，https://doi.org/10.1177/1461444813511926。
46. Lilly Irani, “The Cultural Work of Microwork,” New Media & Society 17, no. 5 (May 2015): 723, https://doi.org/10.1177/1461444813511926.
47. Lilly Irani，《为“数据管理员”伸张正义》公共图书（博客），2015 年 1 月 15 日，http://www.publicbooks.org/justice-for-data-janitors/。
47. Lilly Irani, “Justice for ‘Data Janitors,’ ” Public Books (blog), January 15, 2015, http://www.publicbooks.org/justice-for-data-janitors/.
48. Bin Yu，“数理统计研究所 | IMS 校长致辞：让我们拥有数据科学”，2014 年 7 月，https://imstat.org/2014/10/01/ims-presidential-address-let-us-own-data-science/。
48. Bin Yu, “Institute of Mathematical Statistics | IMS Presidential Address: Let Us Own Data Science,” July 2014, https://imstat.org/2014/10/01/ims -presidential-address-let-us-own-data-science/.
49. Richard Olshen 和 Leo Breiman，“与 Leo Breiman 的对话”，《统计科学》，2001 年，第 196 页。
49. Richard Olshen and Leo Breiman, “A Conversation with Leo Breiman,” Statistical Science, 2001, 196.
50. Olshen 和 Breiman，188。
50. Olshen and Breiman, 188.
51. Leo Breiman，“《统计学的未来报告》：评论”，《统计科学》第 19 卷第 3 期（2004 年）：第 411-411 页。
51. Leo Breiman, “[A Report on the Future of Statistics]: Comment,” Statistical Science 19, no. 3 (2004): 411–411.
52. Leo Breiman，“统计建模：两种文化”，《统计科学》 16 卷，第 3 期（2001 年）：第 201 页。
52. Leo Breiman, “Statistical Modeling: The Two Cultures,” Statistical Science 16, no. 3 (2001): 201.
53. Chambers，“更多或更少的统计数据”，第182页。
53. Chambers, “Greater or Lesser Statistics,” 182.
54. David Madigan 和 Werner Stuetzle，“《统计学的未来报告》：评论”，《统计科学》第 19 卷第 3 期（2004 年）：第 408 页。
54. David Madigan and Werner Stuetzle, “[A Report on the Future of Statistics]: Comment,” Statistical Science 19, no. 3 (2004): 408.
55. 请参阅被广泛引用的麦肯锡报告 Nicolaus Henke 和 Jacques Bughin，“分析时代：在数据驱动的世界中竞争”（麦肯锡全球研究院，2016 年 12 月）。
55. See the much-cited McKinsey report Nicolaus Henke and Jacques Bughin, “The Age of Analytics: Competing in a Data-Driven World” (McKinsey Global Institute, December 2016).
56. Gina Neff 等人，“批评与贡献：改进关键数据研究和数据科学的实践型框架”，大数据5，第 2 期（2017 年 6 月）：85–97，https://doi.org/10.1089/big.2016.0050。
56. Gina Neff et al., “Critique and Contribute: A Practice-Based Framework for Improving Critical Data Studies and Data Science,” Big Data 5, no. 2 (June 2017): 85–97, https://doi.org/10.1089/big.2016.0050.
57. Jennifer Bryan 和 Hadley Wickham，“数据科学：三环马戏团还是大帐篷？”，《计算与图形统计杂志》第 26 卷，第 4 期（2017 年 10 月 2 日）：784–85，https://doi.org/10.1080/10618600.2017.1389743。
57. Jennifer Bryan and Hadley Wickham, “Data Science: A Three Ring Circus or a Big Tent?,” Journal of Computational and Graphical Statistics 26, no. 4 (October 2, 2017): 784–85, https://doi.org/10.1080/10618600.2017.1389743.
58. “Facebook 的数据科学家真的是数据分析师吗”，2017 年 8 月 25 日，https://www.reddit.com/r/datascience/comments/6vv7u2/are_data_scientists_at_facebook_really_data/
58. “Are Data Scientists at Facebook Really Data Analysts,” 25.8.2017, https://www.reddit.com/r/datascience/comments/6vv7u2/are_data_scientists_at_facebook_really_data/
59. Ryan Tibshirani，“德尔福的 COVIDcast 项目：构建用于追踪和预测疫情的数字生态系统的经验教训”，https://docs.google.com/presentation/d/1t_T8BRIkvC5CDOgE4_1PekPw-SThN2 nMJTdieYgdnr4。
59. Ryan Tibshirani, “Delphi’s COVIDcast Project: Lessons from Building a Digital Ecosystem for Tracking and Forecasting the Pandemic,” https://docs.google.com/presentation/d/1t_T8BRIkvC5CDOgE4_1PekPw-SThN2 nMJTdieYgdnr4.
60. Blaise Aguera y Arcas、Margaret Mitchell 和 Alexander Todorov，《面相学的新衣》，Medium（博客），2017 年 5 月 20 日，https://medium.com/@blaisea/physiognomys-new-clothes-f2d4b59fdd6a。
60. Blaise Aguera y Arcas, Margaret Mitchell, and Alexander Todorov, “Physiognomy’s New Clothes,” Medium (blog), May 20, 2017, https://medium.com/@blaisea/physiognomys-new-clothes-f2d4b59fdd6a.
61. Luke Stark 和 Jevan Hutson，《面相人工智能》，《福特汉姆知识产权、媒体与娱乐法杂志》，即将出版，https://doi.org/10.2139/ssrn.3927300，第 5 页（已删除内部引用）。
61. Luke Stark and Jevan Hutson, “Physiognomic Artificial Intelligence,” Ford-ham Intellectual Property, Media & Entertainment Law Journal, no. forthcoming, https://doi.org/10.2139/ssrn.3927300, p. 5 (internal citations removed).
62. D'Ignazio 和 Klein，数据女权主义，特别是。 ch. 2.
62. D’Ignazio and Klein, Data Feminism, esp. ch. 2.
63. 布林致邮件列表 1997 年 10 月 11 日。斯坦福大学 MIDAS 小组的网页保存在 http://infolab.stanford.edu/midas/；数据挖掘小组的邮件列表部分存档在被删除前可在雅虎电子群组上找到；琼斯有部分副本。布林计划复制旧消息，但未能成功。
63. Brin to listserv 10.11.97. The webpage for the Stanford group MIDAS is preserved at http://infolab.stanford.edu/midas/; a partial archive of the listserv of the data mining group was available on Yahoo e-groups before it was deleted; Jones has a partial copy. Brin planned to copy old messages but failed to do so.
第 11 章：数据伦理之战
CHAPTER 11: THE BATTLE FOR DATA ETHICS
1. AI Now Institute，紧缩、不平等和自动化 | AI Now 2018 研讨会，2018 年，https://www.youtube.com/watch?v = gI1KxTrPDLo。
1. AI Now Institute, Austerity, Inequality, and Automation | AI Now 2018 Symposium, 2018, https://www.youtube.com/watch?v = gI1KxTrPDLo.
2. Joy Buolamwini 和 Timnit Gebru，《性别阴影：商业性别分类中的交叉准确性差异》，载于《第一届公平、问责和透明度会议论文集》（公平、问责和透明度会议，PMLR，2018 年），第 77-91 页，https:// proceedings.mlr.press/v81/buolamwini18a.html。另请参阅该项目的配套网站，http://gendershades.org/。
2. Joy Buolamwini and Timnit Gebru, “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification,” in Proceedings of the 1st Conference on Fairness, Accountability and Transparency (Conference on Fairness, Accountability and Transparency, PMLR, 2018), 77–91, https:// proceedings.mlr.press/v81/buolamwini18a.html. See also accompanying website for the project, http://gendershades.org/.
3. Margaret Mitchell 等人，“模型报告的模型卡”，《公平、问责和透明度会议论文集》，FAT* '19（纽约：计算机协会，2019 年），220–29，https://doi.org/10.1145/3287560.3287596。
3. Margaret Mitchell et al., “Model Cards for Model Reporting,” in Proceedings of the Conference on Fairness, Accountability, and Transparency, FAT* ‘19 (New York: Association for Computing Machinery, 2019), 220–29, https://doi.org/10.1145/3287560.3287596.
4. Tom Simonite，“谷歌愿意帮助他人解决人工智能的棘手伦理问题”，《连线》，2021 年 8 月 24 日访问，https://www.wired.com/story/google-help-others-tricky-ethics-ai/。
4. Tom Simonite, “Google Offers to Help Others With the Tricky Ethics of AI,” Wired, accessed August 24, 2021, https://www.wired.com/story/google -help-others-tricky-ethics-ai/.
5. Cade Metz 和 Daisuke Wakabayashi，“谷歌研究员称因发表强调人工智能偏见的论文被解雇”，《纽约时报》，2020 年 12 月 3 日，sec. Technology，https://www.nytimes.com/2020/12/03/technology/google-researcher-timnit-gebru.html。
5. Cade Metz and Daisuke Wakabayashi, “Google Researcher Says She Was Fired Over Paper Highlighting Bias in A.I.,” New York Times, December 3, 2020, sec. Technology, https://www.nytimes.com/2020/12/03/technology /google-researcher-timnit-gebru.html.
6. Adam DI Kramer、Jamie E. Guillory 和 Jeffrey T. Hancock，“社交网络大规模情绪感染的实验证据”，《美国国家科学院院刊》第 111 卷，第 24 期（2014 年 6 月 17 日）：8788–90，https://doi.org/10.1073/pnas.1320040111。
6. Adam D. I. Kramer, Jamie E. Guillory, and Jeffrey T. Hancock, “Experimental Evidence of Massive-Scale Emotional Contagion through Social Networks,” Proceedings of the National Academy of Sciences 111, no. 24 (June 17, 2014): 8788–90, https://doi.org/10.1073/pnas.1320040111.
7. Alex Hern，《Facebook 故意让人们伤心。这应该是最后一根稻草》，《卫报》，2014 年 6 月 30 日，sec。观点，https://www.theguardian.com/commentisfree/2014/jun/30/facebook-sad-manipulating-emotions-socially-responsible-company。
7. Alex Hern, “Facebook Deliberately Made People Sad. This Ought to Be the Final Straw,” The Guardian, June 30, 2014, sec. Opinion, https://www .theguardian.com/commentisfree/2014/jun/30/facebook-sad-manipulating -emotions-socially-responsible-company.
8. Matt Murray，“用户对 Facebook 情绪操纵研究感到愤怒”，TODAY.com，2014 年 6 月 30 日，http://www.today.com/health/users-angered-facebook-emotion-manipulation-study-1D79863 049.Murray。
8. Matt Murray, “Users Angered at Facebook Emotion-Manipulation Study,” TODAY.com, June 30, 2014, http://www.today.com/health/users-angered -facebook-emotion-manipulation-study-1D79863 049.Murray.
9. MJ Salganik，《点点滴滴：数字时代的社会研究》（新泽西州普林斯顿：普林斯顿大学出版社，2017 年），第 282 页。
9. M. J. Salganik, Bit by Bit: Social Research in the Digital Age (Princeton, NJ: Princeton University Press, 2017), 282.
10. Chris Chambers，“Facebook 惨败：康奈尔大学的‘情绪感染’研究是否违反了道德规范？” 《卫报》，2014 年 7 月 1 日，https://www.theguardian.com/science/head-quarters/2014/jul/01/facebook-cornell-study-emotional-contagion-ethics-breach。
10. Chris Chambers, “Facebook Fiasco: Was Cornell’s Study of ‘Emotional Contagion’ an Ethics Breach?,” The Guardian, July 1, 2014, https://www .theguardian.com/science/head-quarters/2014/jul/01/facebook-cornell -study-emotional-contagion-ethics-breach.
11. Allan M. Brandt，“种族主义与研究：塔斯基吉梅毒研究案例”，黑斯廷斯中心报告8，第 6 期（1978 年）：21–29，https://doi.org/10.2307/3561468。
11. Allan M. Brandt, “Racism and Research: The Case of the Tuskegee Syphilis Study,” Hastings Center Report 8, no. 6 (1978): 21–29, https://doi.org/10 .2307/3561468.
12. RA Vonderlehr 等人，“男性黑人中未治疗的梅毒：经治疗和未治疗病例的比较研究”，《美国医学会杂志》 107，第 11 期（1936 年 9 月 12 日）：856–60，https://doi.org/10.1001/jama.1936.02770370020006。
12. R. A. Vonderlehr et al., “Untreated Syphilis in the Male Negro: A Comparative Study of Treated and Untreated Cases,” Journal of the American Medical Association 107, no. 11 (September 12, 1936): 856–60, https://doi.org/10.1001/jama.1936.02770370020006.
13. Susan Reverby，《审视塔斯基吉：臭名昭著的梅毒研究及其遗产》（教堂山：北卡罗来纳大学出版社，2009 年）。
13. Susan Reverby, Examining Tuskegee: The Infamous Syphilis Study and Its Legacy (Chapel Hill: University of North Carolina Press, 2009).
14. 参见 Albert Jonsen，《口述历史》，Bernard Schwetz 访谈，2004 年 5 月 14 日，https://www.hhs.gov/ohrp/education-and-outreach/luminaries-lecture-series/belmont-report-25th-anniversary-interview-ajonsen/index.html。
14. See Albert Jonsen, Oral History, interview by Bernard Schwetz, May 14, 2004, https://www.hhs.gov/ohrp/education-and-outreach/luminaries -lecture-series/belmont-report-25th-anniversary-interview-ajonsen/index.html.
15. 美国国家保护人类受试者和生物医学与行为研究委员会，“贝尔蒙特报告：涉及人类受试者研究的伦理原则和指南”（卫生、教育和福利部，1979 年 4 月 18 日），https://www.hhs.gov/ohrp/sites/default/files/the-belmont-report-508c_FINAL.pdf。
15. National Commission for the Protection of Human Subjects of and Biomedical and Behavioral Research, “The Belmont Report: Ethical Principles & Guidelines for Research Involving Human Subjects” (Department of Health, Education, and Welfare, April 18, 1979), https://www.hhs.gov/ohrp/sites/default/files/the-belmont-report-508c_FINAL.pdf.
16. 报告本身附有上千页的附录，详细阐述了他们关于如何将道德和社会规范付诸实施并成为政府批准的程序规范的想法。
16. The report itself is accompanied by more than a thousand pages of appendices, explaining in detail their thinking on how to operationalize ethics and social norms in a way that could become a government-approved procedural specification.
17. 汤姆·L·博尚普（Tom L. Beauchamp），《坚守原则：论文集》（纽约：牛津大学出版社，2010 年），第 6 页。
17. Tom L. Beauchamp, Standing on Principles: Collected Essays (New York: Oxford University Press, 2010), 6.
18. Karen Lebacqz，LeRoy Walters 采访，2004 年 10 月 26 日，https://www.hhs.gov/ohrp/education-and-outreach/luminaries-lecture-series/belmont-report-25th-anniversary-interview-klebacqz/index.html。
18. Karen Lebacqz, interview by LeRoy Walters, October 26, 2004, https:// www.hhs.gov/ohrp/education-and-outreach/luminaries-lecture-series/ belmont-report-25th-anniversary-interview-klebacqz/index.html.
19. 美国，编辑，报告和建议：机构审查委员会，DHEW 出版物；编号 (OS) 78-0008、78-0009（华盛顿特区：美国卫生、教育与福利部：由文件主管出售，美国政府印刷厂，1978 年）。
19. United States, ed., Report and Recommendations: Institutional Review Boards, DHEW Publication; No. (OS) 78–0008, 78–0009 (Washington, DC: US Department of Health, Education, and Welfare: for sale by the Supt. of Docs., US Govt. Print. Off, 1978).
20.Mike Monteiro，https://muledesign.com/2017/07/a-designers-code-of-ethics。
20. Mike Monteiro, https://muledesign.com/2017/07/a-designers-code-of-ethics.
21. Jacob Metcalf、Emanuel Moss 和 danah boyd，《拥有道德：企业逻辑、硅谷和道德制度化》，《社会研究》 86，第 2 期（2019 年夏季）：449–76。
21. Jacob Metcalf, Emanuel Moss, and danah boyd, “Owning Ethics: Corporate Logics, Silicon Valley, and the Institutionalization of Ethics,” Social Research 86, no. 2 (Summer 2019): 449–76.
22. Brent Mittelstadt，“单靠原则无法保证人工智能合乎道德”，《自然机器智能》1，第 11 期（2019 年 11 月）：501–7，https://doi.org/10.1038/s42256–019–0114–4。
22. Brent Mittelstadt, “Principles Alone Cannot Guarantee Ethical AI,” Nature Machine Intelligence 1, no. 11 (November 2019): 501–7, https://doi.org/10 .1038/s4225 6–019–0114–4.
23. Inioluwa Deborah Raji 等人，“缩小人工智能问责差距：定义内部算法审计的端到端框架”，载于《2020 年公平、问责和透明度会议论文集》，FAT* '20（纽约：计算机协会，2020 年），33–44，https://doi.org/10.1145/3351095.3372873。
23. Inioluwa Deborah Raji et al., “Closing the AI Accountability Gap: Defining an End-to-End Framework for Internal Algorithmic Auditing,” in Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, FAT* ‘20 (New York: Association for Computing Machinery, 2020), 33–44, https://doi.org/10.1145/3351095.3372873.
24. Shannon Vallor，“工程/设计实践的道德工具包”，2018 年 6 月 22 日，https://www.scu.edu/ethics-in-technology-practice/ethical-toolkit/。
24. Shannon Vallor, “An Ethical Toolkit for Engineering/Design Practice,” June 22, 2018, https://www.scu.edu/ethics-in-technology-practice/ethical-toolkit/.
25. Metcalf，Moss 和 boyd，《拥有道德》。
25. Metcalf, Moss, and boyd, “Owning Ethics.”
26. Metcalf、Moss 和 boyd，465。
26. Metcalf, Moss, and boyd, 465.
27. Theodore Vincent Purcell 和 James Weber，《企业伦理制度化：一个案例史》（纽约：总裁协会，美国管理协会首席执行官分部，1979 年），第 6 页；引用自 Ronald R. Sims 的《组织伦理的制度化》，《商业伦理学杂志》第 10 卷，第 7 期（1991 年 7 月 1 日）：第 493 页，https://doi.org/10.1007/BF00383348。
27. Theodore Vincent Purcell and James Weber, Institutionalizing Corporate Ethics: A Case History (New York: Presidents Association, Chief Executive Officers’ Division of American Management Associations, 1979), 6; quoted in Ronald R. Sims, “The Institutionalization of Organizational Ethics,” Journal of Business Ethics 10, no. 7 (July 1, 1991): 493, https://doi.org/10 .1007/BF00383348.
28. Eric Johnson，“AI 将如何改变你的生活？AI Now Institute 创始人 Kate Crawford 和 Meredith Whittaker 解释”，Vox，2019 年 4 月 8 日，https://www.vox.com/podcasts/2019/4/8/18299736/artificial-intelligence-ai-meredith-whittaker-kate-crawford-kara-swisher-decode-podcast -interview。
28. Eric Johnson, “How Will AI Change Your Life? AI Now Institute Founders Kate Crawford and Meredith Whittaker Explain,” Vox, April 8, 2019, https://www.vox.com/podcasts/2019/4/8/1829 973 6/artificial-intelligence -ai-meredith-whittaker-kate-crawford-kara-swisher-decode-podcast -interview.
29. Ben Wagner，“道德作为逃避监管的途径：从‘道德洗白’到道德购物？”，《我思故我在：对欧洲公民的十年剖析》，Emre Bayamlioglu 等主编（阿姆斯特丹大学出版社，2018 年），第 84–89 页，https://doi.org/10.2307/jxtvhrd092.18。
29. Ben Wagner, “Ethics as an Escape from Regulation: From ‘Ethics-Washing’ to Ethics-Shopping?,” in Cogitas Ergo Sum: 10 Years of Profiling the European Citizen, ed. Emre Bayamlioglu et al. (Amsterdam University Press, 2018), 84–89, https://doi.org/10.2307/jxtvhrd092.18.
30. Metcalf，Moss 和 boyd，《拥有道德》。
30. Metcalf, Moss, and boyd, “Owning Ethics.”
31. Henry T. Greely，“大规模基因组生物库的伦理和法律基础令人不安”，《基因组学和人类遗传学年度评论》第 8 卷，第 1 期（2007 年）：343–64，https://doi.org/10.1146/annurev.genom.L080505.115721。
31. Henry T. Greely, “The Uneasy Ethical and Legal Underpinnings of Large-Scale Genomic Biobanks,” Annual Review of Genomics and Human Genetics 8, no. 1 (2007): 343–64, https://doi.org/10.1146/annurev.genom.L080505 .115721.
32. Arvind Narayanan 和 Vitaly Shmatikov，“如何打破 Netflix 奖金数据集的匿名性”（arXiv，2007 年 11 月 22 日），https://doi.org/10.48550/arXiv.cs/0610105。
32. Arvind Narayanan and Vitaly Shmatikov, “How to Break Anonymity of the Netflix Prize Dataset” (arXiv, November 22, 2007), https://doi.org/10 .48550/arXiv.cs/0610105.
33. Pierangela Samarati 和 Latanya Sweeney，“披露信息时保护隐私：K-匿名性及其通过泛化和压制的实施”，1998 年。
33. Pierangela Samarati and Latanya Sweeney, “Protecting Privacy When Disclosing Information: K-Anonymity and Its Enforcement through Generalization and Suppression,” 1998.
34. Cynthia Dwork 和 Moni Naor，“论统计数据库中防止泄露的困难或差异隐私的案例”，《隐私与保密杂志》第 2 卷，第 1 期（2010 年 9 月 1 日）：第 94 页，https://doi.org/10.29012/jpc.v2i1.585。
34. Cynthia Dwork and Moni Naor, “On the Difficulties of Disclosure Prevention in Statistical Databases or The Case for Differential Privacy,” Journal of Privacy and Confidentiality 2, no. 1 (September 1, 2010): 94, https://doi .org/10.29012/jpc.v2i1.585.
35. Cynthia Dwork，《差异隐私》，载于《自动机、语言和编程》，Michele Bugliesi 等编辑，《计算机科学讲稿》（柏林、海德堡：Springer，2006 年），第 4 页，https://doi.org/10.1007/11787006_1。
35. Cynthia Dwork, “Differential Privacy,” in Automata, Languages and Programming, ed. Michele Bugliesi et al., Lecture Notes in Computer Science (Berlin, Heidelberg: Springer, 2006), 4, https://doi.org/10.1007/11787006_1.
36. Cathy O’Neil，《数学毁灭性武器：大数据如何加剧不平等并威胁民主》（纽约：Crown，2016 年）；Virginia Eubanks，《自动化不平等：高科技工具如何剖析、监管和惩罚穷人》（纽约：圣马丁出版社，2017 年）；Ruha Benjamin，《科技之后的种族：新吉姆法典的废奴主义工具》（英国剑桥；马萨诸塞州梅德福：Polity Press，2019 年）。
36. Cathy O’Neil, Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy (New York: Crown, 2016); Virginia Eubanks, Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor (New York: St. Martin’s Press, 2017); Ruha Benjamin, Race after Technology: Abolitionist Tools for the New Jim Code (Cambridge, UK; Medford, MA: Polity Press, 2019).
37. Julia Angwin、Jeff Larson、Surya Mattu 和 Lauren Kirchner，《机器偏见》，ProPublica，2016 年 5 月 23 日，https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing。
37. Julia Angwin, Jeff Larson, Surya Mattu, Lauren Kirchner, “Machine Bias,” ProPublica, May 23, 2016, https://www.propublica.org/article/machine -bias-risk-assessments-in-criminal-sentencing.
38. Arvind Narayanan，教程：21 个公平定义及其政治，2018 年，https://www.youtube.com/watch?v=jIXIuYdnyyk。请参阅 Shubham Jain 的注释，“TL;DS - 21 个公平定义及其政治，作者：Arvind Narayanan，”2019 年 7 月 19 日，https://shubhamjain0594.github.io/post/tlds-arvind-fairness-definitions/。
38. Arvind Narayanan, Tutorial: 21 Fairness Definitions and Their Politics, 2018, https://www.youtube.com/watch?v=jIXIuYdnyyk. See the notes at Shubham Jain, “TL;DS - 21 Fairness Definition and Their Politics by Arvind Narayanan,” July 19, 2019, https://shubhamjain0594.github.io/post/tlds-arvind-fairness-definitions/.
39. Julie Zhuo，“如何设定指标？” 《The Year of the Looking Glass》（博客），2017 年 8 月 10 日，https://medium.com/the-year-of-the-looking-glass/how-do-you-set-metrics-59f78fea7e44。
39. Julie Zhuo, “How Do You Set Metrics?,” The Year of the Looking Glass (blog), August 10, 2017, https://medium.com/the-year-of-the-looking-glass/how-do-you-set-metrics-59f78fea7e44.
40. Michael Kearns 和 Aaron Roth，《道德算法：社会意识算法设计的科学》（纽约：牛津大学出版社，2020 年），第 78 页。
40. Michael Kearns and Aaron Roth, The Ethical Algorithm: The Science of Socially Aware Algorithm Design (New York: Oxford University Press, 2020), 78.
41. Will Douglas Heaven，“预测性警务算法是种族主义的。它们需要被废除”，《麻省理工技术评论》，2020 年 7 月 17 日，https://www.technologyreview.com/2020/07/17/1005396/predictive-policing-algorithms-racist-dismantled-machine-learning-bias-criminal-justice/。
41. Will Douglas Heaven, “Predictive Policing Algorithms Are Racist. They Need to Be Dismantled,” MIT Technology Review, July 17, 2020, https://www.technologyreview.com/2020/07/17/1005396/predictive-policing -algorithms-racist-dismantled-machine-learning-bias-criminal-justice/.
42.Kearns 和Roth，《道德算法》，63页。
42. Kearns and Roth, The Ethical Algorithm, 63.
43. Catherine D'Ignazio 和 Lauren F. Klein，《数据女权主义》（马萨诸塞州剑桥：麻省理工学院出版社，2020 年），61。
43. Catherine D’Ignazio and Lauren F. Klein, Data Feminism (Cambridge, MA: MIT Press, 2020), 61.
44. Matthew Le Bui 和 Safiya Umoja Noble，“我们缺少人工智能的正义道德框架”，载《牛津人工智能伦理手册》（纽约：牛津大学出版社，2020 年），第 178 页，https://doi.org/10.1093/oxfordhb/9780190067397.013.9。
44. Matthew Le Bui and Safiya Umoja Noble, “We’re Missing a Moral Framework of Justice in Artificial Intelligence,” in The Oxford Handbook of Ethics of AI (New York: Oxford University Press, 2020), 178, https://doi.org/10 .1093/oxfordhb/9780190067397.013.9.
45. Julia Powles 和 Helen Nissenbaum，“人工智能中‘解决’偏见的诱惑性转移”，Medium，2018 年 12 月 7 日，https://medium.com/s/story/the-seductive-diversion-of-solving-bias-in-artificial-intelligence-890df5e5ef53。
45. Julia Powles and Helen Nissenbaum, “The Seductive Diversion of ‘Solving’ Bias in Artificial Intelligence,” Medium, December 7, 2018, https:// medium.com/s/story/the-seductive-diversion-of-solving-bias-in-artificial -intelligence-89 0df5e5ef53.
46. Frank Pasquale，“算法问责的第二波浪潮”，《法律与政治经济学》（博客），2019 年 11 月 25 日，https://lpeblog.org/2019/11/25/the-second-wave-of-algorithmic-accountability/。
46. Frank Pasquale, “The Second Wave of Algorithmic Accountability,” Law and Political Economy (blog), November 25, 2019, https://lpeblog.org/2019/11/25/the-second-wave-of-algorithmic-accountability/.
47. Rodrigo Ochigame，“‘道德人工智能’的发明：大型科技公司如何操纵学术界以避免监管”，The Intercept（博客），2019 年 12 月 20 日，https://theintercept.com/2019/12/20/mit-ethical-ai-artificial-intelligence/。
47. Rodrigo Ochigame, “The Invention of ‘Ethical AI’: How Big Tech Manipulates Academia to Avoid Regulation,” The Intercept (blog), December 20, 2019, https://theintercept.com/2019/12/20/mit-ethical-ai-artificial-intelligence/.
48. Thao Phan 等人，“美德经济：‘伦理’在大型科技公司中的流通”，《科学作为文化》，2021 年 11 月 4 日，第 7 页，https://doi.org/10.1080/09505431.2021.1990875。
48. Thao Phan et al., “Economies of Virtue: The Circulation of ‘Ethics’ in Big Tech,” Science as Culture, November 4, 2021, 7, https://doi.org/10.1080/09505431.2021.1990875.
49. Shoshana Zuboff，《监控资本主义时代：在权力新前沿为人类未来而战》（纽约：PublicAffairs，2019 年）。
49. Shoshana Zuboff, The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power (New York: PublicAffairs, 2019).
第 12 章：说服、广告和风险投资
CHAPTER 12: PERSUASION, ADS, AND VENTURE CAPITAL
1. Herbert Simon，《为信息丰富的世界设计组织》，载于《计算机、通信和公共利益》，Martin Greenberger 主编（巴尔的摩：约翰霍普金斯出版社，1971 年），第 40 页。
1. Herbert Simon, “Designing Organizations for an Information-Rich World,” in Computers, Communications, and the Public Interest, ed. Martin Greenberger (Baltimore: Johns Hopkins Press, 1971), 40.
2. Paul Lewis，《虚构超越现实：Youtube 算法如何扭曲真相》，《卫报》，2018 年 2 月 2 日，第 36 版，技术，http://www.theguardian.com/technology/2018/feb/02/how-youtubes-algorithm-distorts-truth。
2. Paul Lewis, “Fiction Is Outperforming Reality: How Youtube’s Algorithm Distorts Truth,” The Guardian, February 2, 2018, sec. Technology, http://www.theguardian.com/technology/2018/feb/02/how-youtubes-algorithm-distorts-truth.
3. “复活节的阳光在现代游行的阴影中找到了过去的影子”，《纽约时报》，1929 年 4 月 1 日，https://timesmachine.nytimes.com/timesmachine/1929/04/01/95899706.pdf。
3. “Easter Sun Finds the Past in Shadow at Modern Parade,” New York Times, April 1, 1929, https://timesmachine.nytimes.com/timesmachine/1929/04/01/95899706.pdf.
4.爱德华·伯内斯（Edward L. Bernays），《宣传》（纽约：H. Liveright，1928 年）。
4. Edward L. Bernays, Propaganda (New York: H. Liveright, 1928).
5. Simon，“为信息丰富的世界设计组织”，第41页。
5. Simon, “Designing Organizations for an Information-Rich World,” 41.
6.理查德·塞拉，《电视传递人们（1973年）》，《通讯》 48卷，第1期（1988年）：42–44页。
6. Richard Serra, “TV Delivers People (1973),” Communications 48, no. 1 (1988): 42–44.
7.尼尔·波兹曼，《娱乐至死：娱乐时代的公共话语》（纽约：企鹅图书，1986 年）。
7. Neil Postman, Amusing Ourselves to Death: Public Discourse in the Age of Show Business (New York: Penguin Books, 1986).
8.如需更深入地了解波兹曼及其媒体决定论，请参阅 Siva Vaidhyanathan 的《反社会媒体：Facebook 如何切断我们之间的联系并破坏民主》（纽约：牛津大学出版社，2018 年），第 21-26 页。
8. For a deeper engagement with Postman and his forms of media determinism, see Siva Vaidhyanathan, Antisocial Media: How Facebook Disconnects Us and Undermines Democracy (New York: Oxford University Press, 2018), pp. 21–26.
9. https://www.w3.org/People/Berners-Lee/1991/08/art-6484.txt。
9. https://www.w3.org/People/Berners-Lee/1991/08/art-6484.txt.
10. Ethan Zuckerman，《互联网的原罪》，《大西洋月刊》，2014 年 8 月 14 日，https://www.theatlantic.com/technology/archive/2014/08/advertising-is-the-internets-original-sin/376041/。
10. Ethan Zuckerman, “The Internet’s Original Sin,” The Atlantic, August 14, 2014, https://www.theatlantic.com/technology/archive/2014/08/advertising -is-the-internets-original-sin/376041/.
11. Michael H. Goldhaber，《注意力经济与网络》，《First Monday 》第2卷，第4期（1997年4月7日），http://firstmonday.org/ojs/index.php/fm/article /view/519。
11. Michael H. Goldhaber, “The Attention Economy and the Net,” First Monday 2, no. 4 (April 7, 1997), http://firstmonday.org/ojs/index.php/fm/article /view/519.
12. Goldhaber，“注意力经济与网络”。
12. Goldhaber, “The Attention Economy and the Net.”
13. Daniel Thomas 和 Shannon Bond，“BuzzFeed 老板发现移动端社交内容的自然契合点”，《金融时报》，2016 年 3 月 14 日，https://www.ft.com/content/4f661ea8-e782–11e5-a09b-1f8b0d268c39。
13. Daniel Thomas and Shannon Bond, “BuzzFeed Boss Finds Natural Fit for Social Content on Mobile,” Financial Times, March 14, 2016, https://www .ft.com/content/4f661ea8-e782–11e5-a09b-1f8b0d268c39.
14. 戈德哈伯的预言并非全部都已成真：他还预言了一种纯粹以注意力为基础的经济，金钱在其中将不起作用。尽管现在人们可以交换手机通话时间，或与另一位学习互补语言的学生交换语言课程的时间，但我们还没有实现注意力完全取代金钱的经济。
14. Not all the prophecies of Goldhaber have yet come to pass: he also prophesied a purely attention-based economy in which money would play no role. Although one can now barter mobile-phone minutes, or trade time in language lessons with another student in complementary languages, we do not yet have economies in which attention has fully replaced money.
15. Goldhaber，“注意力经济与网络”。
15. Goldhaber, “The Attention Economy and the Net.”
16. Stewart Brand，《世界信息经济》， 《全球概览》，冬季号（1986 年）：88。
16. Stewart Brand, “The World Information Economy,” The Whole Earth Catalog, no. Winter (1986): 88.
17. 有关简要概述，请参阅 Christina Spurgeon，《在线广告》，《劳特利奇全球互联网历史指南》（劳特利奇，2017 年）；Joseph Turow，《每日的你：新广告行业如何定义你的身份和价值》（康涅狄格州纽黑文：耶鲁大学出版社，2011 年）。
17. For a concise overview, see Christina Spurgeon, “Online Advertising,” in The Routledge Companion to Global Internet Histories (Routledge, 2017); Joseph Turow, The Daily You: How the New Advertising Industry Is Defining Your Identity and Your Worth (New Haven, CT: Yale University Press, 2011).
18. 请参阅 Tim O'Reilly，《什么是 Web 2.0》，2005 年 9 月 30 日，https://www.oreilly.com/pub/a//web2/archive/what-is-web-20.html。有关对这一转变的新颖性的质疑，请参阅 Matthew Allen，《什么是 Web 2.0？版本作为互联网历史的主导模式》，《新媒体与社会》第 15 卷第 2 期（2013 年 3 月 1 日）：260–75，https://doi.org/10.1177/1461444812451567。
18. See Tim O’Reilly, “What Is Web 2.0,” September 30, 2005, https:// www.oreilly.com/pub/a//web2/archive/what-is-web-20.html. For skepticism about the novelty of this shift, see Matthew Allen, “What Was Web 2.0? Versions as the Dominant Mode of Internet History,” New Media & Society 15, no. 2 (March 1, 2013): 260–75, https://doi.org/10 .1177/1461444812451567.
19. Nick Couldry 和 Joseph Turow，“大数据、大问题 | 广告、大数据和公共领域的清理：营销人员对内容补贴的新方法”，《国际传播学杂志》第 8 卷（2014 年 6 月 16 日）：1714。
19. Nick Couldry and Joseph Turow, “Big Data, Big Questions | Advertising, Big Data and the Clearance of the Public Realm: Marketers’ New Approaches to the Content Subsidy,” International Journal of Communication 8 (June 16, 2014): 1714.
20. Kim Cleland，“媒体购买与规划：营销人员需要有关互联网广告购买价值的可靠数据：对媒体选项比较信息的需求不断增长”，《广告时代》 ，1998 年 8 月 3 日，第 18 页；Turow 的《每日你》第 61 页对此进行了讨论。
20. Kim Cleland, “Media Buying & Planning: Marketers Want Solid Data on Value of Internet Ad Buys: Demand Swells for Information That Compares Media Options,” Advertising Age, August 3, 1998, S18; discussed in Turow, The Daily You, 61.
21. Cleland，“营销人员想要可靠的数据”；Turow 在《每日你》第 61 页中对此进行了讨论。
21. Cleland, “Marketers Want Solid Data”; discussed in Turow, The Daily You, 61.
22. Rick Bruner，“‘Cookie’提案可能阻碍在线广告：隐私支持者推动加强数据控制”，《广告时代》，1997 年 3 月 16 日，第 16 页；Turow 的《每日你》第 58 页对此进行了讨论。
22. Rick Bruner, “ ‘Cookie’ Proposal Could Hinder Online Advertising: Privacy Backers Push for More Data Controls,” Advertising Age, March 16, 1997, 16; discussed in Turow, The Daily You, 58.
23. 引自 Meg Leta Jones 的《Cookies：争议的遗产》，《互联网历史》第 4 卷，第 1 期（2020 年 1 月 2 日）：第 94 页，https://doi.org/10.1080/24701475 .2020.1725852。另请参阅 David M. Kristol 的《HTTP Cookies：标准、隐私和政治》，《ACM 互联网技术交易》第 1 卷，第 2 期（2001 年 11 月 1 日）：第 151–98 页，https://doi.org/10.1145/502152.502153。
23. Quoted in Meg Leta Jones, “Cookies: A Legacy of Controversy,” Internet Histories 4, no. 1 (January 2, 2020): 94, https://doi.org/10.1080/24701475 .2020.1725852. See also David M. Kristol, “HTTP Cookies: Standards, Privacy, and Politics,” ACM Transactions on Internet Technology 1, no. 2 (November 1, 2001): 151–98, https://doi.org/10.1145/502152.502153.
24. CNET 新闻工作人员，“广告在数量中寻找权力”，CNET，1996 年 11 月 4 日，https://www.cnet.com/tech/tech-industry/ads-find-strength-in-numbers/。
24. CNET News staff, “Ads Find Strength in Numbers,” CNET, November 4, 1996, https://www.cnet.com/tech/tech-industry/ads-find-strength-in -numbers/.
25. 参见 Jones，《Cookies》，第 95 页；Matthew Crain，《利润高于隐私：监控广告如何征服互联网》（明尼阿波利斯：明尼苏达大学出版社，2021 年），第 125–29 页。
25. See Jones, “Cookies,” 95; Matthew Crain, Profit Over Privacy: How Surveillance Advertising Conquered the Internet (Minneapolis: University of Minnesota Press, 2021), 125–29.
26.Crain ，《利润高于隐私》，129页。
26. Crain, Profit Over Privacy, 129.
27. Susan Wojcicki，“让广告更有趣”，Google 官方博客(blog)，2009 年 3 月 11 日，https://googleblog.blogspot.com/2009/03/making-ads-more-interesting.html。
27. Susan Wojcicki, “Making Ads More Interesting,” Official Google Blog (blog), March 11, 2009, https://googleblog.blogspot.com/2009/03/making -ads-more-interesting.html.
28.Crain ，《利润高于隐私》，95页。
28. Crain, Profit Over Privacy, 95.
29. 亚当·德安杰洛，Quora，2010 年，https://www.quora.com/What-was-Adam-DAngelos-biggest-contribution-to-Facebook/answer/Adam-D Angelo。
29. Adam D’Angelo, Quora, 2010, https://www.quora.com/What-was-Adam -DAngelos-biggest-contribution-to-Facebook/answer/Adam-D Angelo.
30. Ashlee Vance，《我们信任广告》，《彭博商业周刊》 ，第 4521 期（2017 年 5 月 8 日）：第 6-7 页。
30. Ashlee Vance, “In Ads We Trust,” Bloomberg Businessweek, no. 4521 (May 8, 2017): 6–7.
31. John White，《网站优化的 Bandit 算法》（O'Reilly Media, Inc.，2012 年）；William R Thompson，“从两个样本的证据来看，一个未知概率超过另一个未知概率的可能性”，《Biometrika》第 25 卷，第 3/4 期（1933 年）：285-94 页。
31. John White, Bandit Algorithms for Website Optimization (O’Reilly Media, Inc., 2012); William R Thompson, “On the Likelihood That One Unknown Probability Exceeds Another in View of the Evidence of Two Samples,” Biometrika 25, no. 3/4 (1933): 285–94.
32. 有关 Facebook 效应的细致入微的描述，拒绝简单的妖魔化叙述，请参阅 Vaidhyanathan 的《反社会媒体》，第 16-17 页。
32. For a nuanced account of Facebook’s effects that rejects a simple demonization narrative, see Vaidhyanathan, Antisocial Media, e.g., at pp. 16–17.
33. James Grimmelmann，“平台即信息”，SSRN 学术论文（纽约州罗切斯特：社会科学研究网络，2018 年 3 月 1 日），https://papers.ssrn.com/abstract=3132758。
33. James Grimmelmann, “The Platform Is the Message,” SSRN Scholarly Paper (Rochester, NY: Social Science Research Network, March 1, 2018), https://papers.ssrn.com/abstract=3132758.
34. Zeynep Tufekci，“工程化公众：大数据、监控和计算政治”，《第一个星期一》，2014 年 7 月 2 日，https://doi.org/10.5210/fm.v19i7.4901。
34. Zeynep Tufekci, “Engineering the Public: Big Data, Surveillance and Computational Politics,” First Monday, July 2, 2014, https://doi.org/10.5210/fm .v19i7.4901.
35. Edward L. Bernays，《同意的工程》，《美国政治和社会科学院年鉴》第 250 卷（1947 年）：第 115 页。
35. Edward L. Bernays, “The Engineering of Consent,” The Annals of the American Academy of Political and Social Science 250 (1947): 115.
36. Salman Haqqi，“奥巴马连任的秘密武器：巴基斯坦科学家 Rayid Ghani”，DAWN.COM，2013 年 1 月 21 日，https://www.dawn.com/2013/01/21/obamas-secret-weapon-in-re-election-pakistani-scientist-rayid-ghani/。
36. Salman Haqqi, “Obama’s Secret Weapon in Re-Election: Pakistani Scientist Rayid Ghani,” DAWN.COM, January 21, 2013, https://www.dawn .com/2013/01/21/obamas-secret-weapon-in-re-election-pakistani-scientist -rayid-ghani/.
37. Rayid Ghani 等人，“个人消费者模型和个性化零售促销的数据挖掘”，数据挖掘方法与应用，2007 年，第 215 页。
37. Rayid Ghani et al., “Data Mining for Individual Consumer Models and Personalized Retail Promotions,” Data Mining Methods and Applications, 2007, 215.
38. Ethan Roeder，《我不是老大哥》，《纽约时报》 ，2012 年 12 月 6 日，http://www.nytimes.com/2012/12/06/opinion/i-am-not-big-brother.html ?_r = 0。
38. Ethan Roeder, “I Am Not Big Brother,” New York Times, December 6, 2012, http://www.nytimes.com/2012/12/06/opinion/i-am-not-big-brother.html ?_r = 0.
39. Zeynep Tufekci，“是的，大型平台可以改变其商业模式”，《连线》，2018 年 12 月 17 日，https://www.wired.com/story/big-platforms-could-change-business-models/。
39. Zeynep Tufekci, “Yes, Big Platforms Could Change Their Business Models,” Wired, December 17, 2018, https://www.wired.com/story/big -platforms-could-change-business-models/.
40. Tufekci，“工程公众”。
40. Tufekci, “Engineering the Public.”
41. MJ Salganik，《点点滴滴：数字时代的社会研究》（新泽西州普林斯顿：普林斯顿大学出版社，2017 年），第 10 页。
41. M. J. Salganik, Bit by Bit: Social Research in the Digital Age (Princeton, NJ: Princeton University Press, 2017), 10.
42. Mike Butcher，《剑桥分析公司首席执行官接受 TechCrunch 采访，谈论特朗普、希拉里和未来》，TechCrunch，2017 年 11 月 6 日，https://social.techcrunch.com/20r7/11/06/cambridge-analytica-ceo-talks-to-techcrunch-about-trump-hilary-and-the-future/。
42. Mike Butcher, “Cambridge Analytica CEO Talks to TechCrunch about Trump, Hillary and the Future,” TechCrunch, November 6, 2017, https://social.techcrunch.com/20r7/11/06/cambridge-analytica-ceo-talks-to-techcrunch-about-trump-hilary-and-the-future/.
43. Trenholme J. Griffin，《企业家的十二堂课》（纽约：哥伦比亚商学院出版社，哥伦比亚大学出版社，2017年），第146页。
43. Trenholme J. Griffin, A DozenLessonsforEntrepreneurs (New York: Columbia Business School Publishing, Columbia University Press, 2017), 146.
44. AnnaLee Saxenian，《区域优势：硅谷和 128 号公路的文化与竞争，作者新作序》（马萨诸塞州剑桥：哈佛大学出版社，1996 年）；Christophe Lécuyer，《打造硅谷：1930-1970 年的创新与高科技发展》（马萨诸塞州剑桥：麻省理工学院出版社，2006 年）。
44. AnnaLee Saxenian, Regional Advantage: Culture and Competition in Silicon Valley and Route 128, With a New Preface by the Author (Cambridge, MA: Harvard University Press, 1996); Christophe Lécuyer, Making Silicon Valley: Innovation and the Growth of High Tech, 1930–1970 (Cambridge, MA: MIT Press, 2006).
45. Josh Lerner，“政府作为风险投资家：SBIR 计划的长期影响”，《私募股权期刊》第 3 卷，第 2 期（2000 年）：第 55-78 页。感谢 Ella Coon 强调这一点。
45. Josh Lerner, “The Government as Venture Capitalist: The Long-Run Impact of the SBIR Program,” The Journal of Private Equity 3, no. 2 (2000): 55–78. Thanks to Ella Coon for stressing this.
46. Jerry Neumann，《热寂：20 世纪 80 年代的风险投资》，Reaction Wheel（博客），2015 年 1 月 8 日，https://reactionwheel.net/2015/01/80s-vc .html。
46. Jerry Neumann, “Heat Death: Venture Capital in the 1980s,” Reaction Wheel (blog), January 8, 2015, https://reactionwheel.net/2015/01/80s-vc .html.
47. 汤姆·尼古拉斯（Tom Nicholas，VC：美国历史）（马萨诸塞州剑桥：哈佛大学出版社，2019 年）。
47. Tom Nicholas, VC: An American History (Cambridge, MA: Harvard University Press, 2019).
48. Katrina Brooker，《WeFail：孙正义与亚当·诺伊曼关系的失败如何让 WeWork 走上灾难之路》，《Fast Company》，2019 年 11 月 15 日，https://www.fastcompany.com/90426446/wefail-how-the-doomed-masa-son-adam-neumann-relationship-set-wework-on-the-road-to-dis aster。
48. Katrina Brooker, “WeFail: How the Doomed Masa Son-Adam Neumann Relationship Set WeWork on the Road to Disaster,” Fast Company, November 15, 2019, https://www.fastcompany.com/90426446/wefail-how-the-doomed -masa-son-adam-neumann-relationship-set-wework-on-the-road-to-dis aster.
49. 李开复，《人工智能超级大国：中国、硅谷和新世界秩序》（波士顿：霍顿·米夫林哈考特，2019 年）。
49. Kai-Fu Lee, AI Superpowers: China, Silicon Valley, and the New World Order (Boston: Houghton Mifflin Harcourt, 2019).
50. Ryan Mac、Charlie Warzel 和 Alex Kantrowitz，“不惜一切代价追求增长：Facebook 高管在 2016 年备忘录中为数据收集辩护 — 并警告称 Facebook 可能害死人”，Buzz Feed News，2018 年 3 月 29 日，https://www.buzzfeednews.com/article/ryanmac/growth-at-any-cost-top-facebook-executive-defended-data。
50. Ryan Mac, Charlie Warzel, and Alex Kantrowitz, “Growth at Any Cost: Top Facebook Executive Defended Data Collection in 2016 Memo—And Warned that Facebook Could Get People Killed,” Buzz Feed News, March 29, 2018, https://www.buzzfeednews.com/article/ryanmac/growth-at-any-cost-top-facebook-executive-defended-data.
51. Paul Lewis，“‘虚构超越现实’：Youtube 算法如何扭曲真相”，《卫报》，2018 年 2 月 2 日，第 36 版，技术，http://www.theguardian.com/technology/2018/feb/02/how-youtubes-algorithm-distorts-truth。
51. Paul Lewis, “ ‘Fiction Is Outperforming Reality’: How Youtube’s Algorithm Distorts Truth,” The Guardian, February 2, 2018, sec. Technology, http://www.theguardian.com/technology/2018/feb/02/how-youtubes-algorithm-distorts-truth.
52. Jane Jacobs，《经济的本质》（纽约：现代图书馆，2000 年）。
52. Jane Jacobs, The Nature of Economies (New York: Modern Library, 2000).
第 13 章：超越解决方案主义的解决方案
CHAPTER 13: SOLUTIONS BEYOND SOLUTIONISM
1. Karl Manheim 和 Lyric Kaplan，《人工智能：对隐私和民主的风险》，《耶鲁法律与技术杂志》第 21 卷第 1 期（2019 年）：第 180、181 页。
1. Karl Manheim and Lyric Kaplan, “Artificial Intelligence: Risks to Privacy and Democracy,” Yale Journal of Law & Technology 21, no. 1 (2019): 180, 181.
2. William H. Janeway，《创新经济中的资本主义：市场、投机和国家》（英国剑桥：剑桥大学出版社，2012 年）。
2. William H. Janeway, Doing Capitalism in the Innovation Economy: Markets, Speculation and the State. (Cambridge, UK: Cambridge University Press, 2012).
3. Marshall Kirkpatrick，“Facebook 的扎克伯格称隐私时代已经结束”，《纽约时报》，2010 年 1 月 10 日，https://archive.nytimes.com/ www.nytimes.com/external/readwriteweb/2010/01/10/10readwriteweb-facebooks-zuckerberg-says-the-age-of-privac-82963.html。
3. Marshall Kirkpatrick, “Facebook’s Zuckerberg Says The Age of Privacy Is Over,” New York Times, January 10, 2010, https://archive.nytimes.com/ www.nytimes.com/external/readwriteweb/2010/01/10/10readwriteweb -facebooks-zuckerberg-says-the-age-of-privac-82963.html.
4.蒂姆·库克，“我们相信隐私是一项基本人权。无论您生活在哪个国家，都应根据四项基本原则保护这项权利”，推文，@tim_cook（博客），2018 年 10 月 24 日，https://twitter.com/tim_cook/status/1055035539915718656。
4. Tim Cook, “We Believe That Privacy Is a Fundamental Human Right. No Matter What Country You Live in, That Right Should Be Protected in Keeping with Four Essential Principles,” Tweet, @tim_cook (blog), October 24, 2018, https://twitter.com/tim_cook/status/1055035539915718656.
5. Blake Lemoine，《谷歌的道德 AI 历史》，Medium，2021 年 5 月 17 日，https://cajundiscordian.medium.com/the-history-of-ethical-ai-at-google-d2f997985233。
5. Blake Lemoine, “The History of Ethical AI at Google,” Medium, May 17, 2021, https://cajundiscordian.medium.com/the-history-of-ethical-ai-at-google-d2f997985233.
6. Urooba Jamal，“一名被谷歌解雇的工程师表示，其 AI 聊天机器人‘相当种族主义’，谷歌的 AI 伦理是一块‘遮羞布’”，Business Insider，2022 年 7 月 31 日，https://www.businessinsider.com/google-engrneer-blake-lemoine-ai-ethics-lamda-racist-2022-7。
6. Urooba Jamal, “An Engineer Who Was Fired by Google Says Its AI Chatbot Is ‘Pretty Racist’ and That AI Ethics at Google Are a ‘Fig Leaf,’ ” Business Insider, July 31, 2022, https://www.businessinsider.com/google-engrneer -blake-lemoine-ai-ethics-lamda-racist-2022–7.
7. danah boyd，《我们在哪里可以找到道德？》，Medium，2016 年 6 月 15 日，https://points.datasociety.net/where-do-we-find-ethics-d0b9e8a7f4e6；引用 Audre Lorde 的《主人的工具永远不会拆除主人的房子》，《局外人姐妹：散文和演讲》（纽约州特鲁曼斯堡：Crossing Press，1984 年），第 110–14 页。
7. danah boyd, “Where Do We Find Ethics?,” Medium, June 15, 2016, https://points.datasociety.net/where-do-we-find-ethics-d0b9e8a7f4e6; citing Audre Lorde, “The Master’s Tools Will Never Dismantle the Master’s House,” in Sister Outsider: Essays and Speeches (Trumansburg, NY: Crossing Press, 1984), 110–14.
8. Anna Kramer，“Twitter 如何聘请科技界最大的批评者来打造合乎道德的人工智能”，《Protocol—科技的人、权力和政治》，2021 年 6 月 23 日，https://www.protocol.com/workplace/twitter-ethical-ai-meta。
8. Anna Kramer, “How Twitter Hired Tech’s Biggest Critics to Build Ethical AI,” Protocol—The people, power and politics of tech, June 23, 2021, https://www.protocol.com/workplace/twitter-ethical-ai-meta.
9. Michael Kearns 和 Aaron Roth，《道德算法：社会意识算法设计的科学》（纽约：牛津大学出版社，2020 年），第 16 页。
9. Michael Kearns and Aaron Roth, The Ethical Algorithm: The Science of Socially Aware Algorithm Design (New York: Oxford University Press, 2020), 16.
10. Kearns 和 Roth，16。
10. Kearns and Roth, 16.
11. Cynthia Rudin，“停止解释高风险决策的黑箱机器学习模型，转而使用可解释的模型”，《自然机器智能》1，第 5 期（2019 年 5 月）：10，https://doi.org/10.1038/s42256 -019–0048-x。
11. Cynthia Rudin, “Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead,” Nature Machine Intelligence 1, no. 5 (May 2019): 10, https://doi.org/10.1038/s42256 -019–0048-x.
12. Annette Zimmermann、Elena Di Rosa 和 Hochan Kim，“技术无法解决算法不公正”，《波士顿评论》，2019 年 12 月 12 日，https://bostonreview.net/science-nature-politics/annette-zimmermann-elena-di-rosa-hochan-kim-technology-cant-fix-algorithmic。
12. Annette Zimmermann, Elena Di Rosa, and Hochan Kim, “Technology Can’t Fix Algorithmic Injustice,” Boston Review, December 12, 2019, https://bostonreview.net/science-nature-politics/annette-zimmermann -elena-di-rosa-hochan-kim-technology-cant-fix-algorithmic.
13. 齐默尔曼、迪罗莎和金。
13. Zimmermann, Di Rosa, and Kim.
14. Gina Neff 等人，“批评与贡献：改进关键数据研究和数据科学的实践型框架”，大数据5，第 2 期（2017 年 6 月）：85–97，https://doi.org/10.1089/big.2016.0050。
14. Gina Neff et al., “Critique and Contribute: A Practice-Based Framework for Improving Critical Data Studies and Data Science,” Big Data 5, no. 2 (June 2017): 85–97, https://doi.org/10.1089/big.2016.0050.
15. Mike Isaac，《超级动力：Uber 之战》（纽约：WW Norton & Company，2020 年），第 16 章。
15. Mike Isaac, Super Pumped: The Battle for Uber (New York: W. W. Norton & Company, 2020), ch. 16.
16. Kate O'Flaherty，“苹果的隐私功能将使 Facebook 损失 120 亿美元”，《福布斯》，2022 年 4 月 23 日，https://www.forbes.com/sites/kateoflahertyuk/2022/04/23/apple-just-issued-stunning-12-billion-blow-to-facebook/。
16. Kate O’Flaherty, “Apple’s Privacy Features Will Cost Facebook $12 Billion,” Forbes, April 23, 2022, https://www.forbes.com/sites/ kateoflahertyuk/2022/04/23/apple-just-issued-stunning-12-billion-blow-to -facebook/.
17. Yochai Benkler 等，“社会动员与网络化公共领域：描绘 SOPA-PIPA 之争”，《政治传播》第 32 卷第 4 期（2015 年 10 月 2 日）：594–624 页，https://doi.org/10.1080/10584609.2014.986349。
17. Yochai Benkler et al., “Social Mobilization and the Networked Public Sphere: Mapping the SOPA-PIPA Debate,” Political Communication 32, no. 4 (October 2, 2015): 594–624, https://doi.org/10.1080/10584609.2014 .986349.
18. 有关持续进行国家监管必要性的有力思考，请参阅弗兰克·帕斯夸莱 (Frank Pasquale) 的《黑箱社会：控制金钱和信息的秘密算法》（马萨诸塞州剑桥：哈佛大学出版社，2015 年）。
18. For a powerful reflection on the continuing need for state regulation, see Frank Pasquale, The Black Box Society: The Secret Algorithms That Control Money and Information (Cambridge, MA: Harvard University Press, 2015).
19. Amy Kapczynski，《信息资本主义法则》，《耶鲁法律杂志》 129，第5期（2020年），1465页。
19. Amy Kapczynski, “The Law of Informational Capitalism,” The Yale Law Journal 129, n. 5 (2020), 1465.
20. Karl Manheim 和 Lyric Kaplan，“人工智能：对隐私和民主的风险”，《耶鲁法律与技术杂志》第 21 卷 (2019 年)：第 162 页。
20. Karl Manheim and Lyric Kaplan, “Artificial Intelligence: Risks to Privacy and Democracy,” Yale Journal of Law & Technology 21 (2019): 162.
21. Michael Kearns 和 Aaron Roth，“道德算法设计应指导技术监管”，布鲁金斯学会（博客），2020 年 1 月 13 日，https://www.brookings.edu/research/ethical-algorithm-design-should-guide-technology-regulation/。
21. Michael Kearns and Aaron Roth, “Ethical Algorithm Design Should Guide Technology Regulation,” Brookings (blog), January 13, 2020, https://www .brookings.edu/research/ethical-algorithm-design-should-guide-technology -regulation/.
22. Morgan Meaker，“Meta 收购 Giphy 失败，结局可能很惨烈科技的消费狂潮”，《连线》，2021 年 12 月 3 日，https://www.wired.com/story/facebook-giphy-cma-global-template/。
22. Morgan Meaker, “Meta’s Failed Giphy Deal Could End Big Tech’s Spending Spree,” Wired, December 3, 2021, https://www.wired.com/story/facebook-giphy-cma-global-template/.
23. Manheim 和 Kaplan，“人工智能：对隐私和民主的风险”，第 186 页。
23. Manheim and Kaplan, “Artificial Intelligence: Risks to Privacy and Democracy,” 186.
24. 转引自 Lina M. Khan，《亚马逊的反垄断悖论》，《耶鲁法律杂志》 ，2017 年，第 740 页。
24. Quoted in Lina M. Khan, “Amazon’s Antitrust Paradox,” The Yale Law Journal, 2017, 740.
25. Patrice Bougette、Marc Deschamps 和 Frédéric Marty，“当经济学遇见反垄断：第二芝加哥学派与反垄断法的经济化”，Enterprise & Society 16，第 2 期（2015 年 6 月）：313–53，https://doi.org/10.1017/eso.2014.18。
25. Patrice Bougette, Marc Deschamps, and Frédéric Marty, “When Economics Met Antitrust: The Second Chicago School and the Economization of Antitrust Law,” Enterprise & Society 16, no. 2 (June 2015): 313–53, https://doi.org/10.1017/eso.2014.18.
26. “通用数据保护条例（GDPR）——官方法律文本”，通用数据保护条例（GDPR），https://gdpr-info.eu/，第 22 条。
26. “General Data Protection Regulation (GDPR)—Official Legal Text,” General Data Protection Regulation (GDPR), https://gdpr-info.eu/, art 22.
27. 梅格·莱塔·琼斯（Meg Leta Jones），《Ctrl + Z：被遗忘的权利》（纽约：纽约大学出版社，2016 年）。
27. Meg Leta Jones, Ctrl + Z: The Right to Be Forgotten (New York: New York University Press, 2016).
28. Khan，“亚马逊的反垄断悖论”。
28. Khan, “Amazon’s Antitrust Paradox.”
29. Sarah T. Roberts，《屏幕背后：社交媒体阴影下的内容审核》（康涅狄格州纽黑文：耶鲁大学出版社，2019 年）。
29. Sarah T. Roberts, Behind the Screen: Content Moderation in the Shadows of Social Media (New Haven, CT: Yale University Press, 2019).
30. Paul M. Barrett，“谁来管理社交媒体巨头？呼吁终止外包”（纽约大学斯特恩商业与人权中心，2020 年 6 月），第 4 页，https://static1.squarespace.com/static/5b6df958f8370af3217d4178/t/5ed9854bf618c710cb55be98/1591313740497/NYU+Content+Moderation+Report_June+8+2020.pdf。
30. Paul M. Barrett, “Who Moderates the Social Media Giants? A Call to End Outsourcing” (NYU Stern—Center for Business and Human Rights, June 2020), 4, https://static1.squarespace.com/static/5b6df958 f8370af3217d4178/t/5ed9854bf618c710cb55be98/1591313740497/NYU + Content+Moderation+Report_June +8+2020.pdf.
31. Jeff Kosseff，《创造互联网的二十六个词》（纽约州伊萨卡：康奈尔大学出版社，2019 年）。
31. Jeff Kosseff, The Twenty-Six Words That Created the Internet (Ithaca, NY: Cornell University Press, 2019).
32. Jennifer S. Fan，“员工作为监管者：高科技公司的新私人秩序”，《犹他法律评论》，2019 年第 5 期（2020 年）：55。
32. Jennifer S. Fan, “Employees as Regulators: The New Private Ordering in High Technology Companies,” Utah Law Review, Vol. 2019, no. 5 (2020): 55.
33. Alexis C. Madrigal，“硅谷筛分：科技行业泄密时间表”，《大西洋月刊》，2018 年 10 月 10 日，https://www.theatlantic.com/technology/archive/2018/10/timeline-tech-industry-leaks/572593/。
33. Alexis C. Madrigal, “Silicon Valley Sieve: A Timeline of Tech-Industry Leaks,” The Atlantic, October 10, 2018, https://www.theatlantic.com/technology/archive/2018/10/timeline-tech-industry-leaks/572593/.
34. Daisuke Wakabayashi，“在谷歌，员工主导的努力发现男性薪酬高于女性”，纽约时报，2017 年 9 月 8 日，https://www.nytimes.com/2017/09/08/technology/google-salaries-gender-disparity .html。
34. Daisuke Wakabayashi, “At Google, Employee-Led Effort Finds Men Are Paid More Than Women,” New York Times, September 8, 2017, https:// www.nytimes.com/2017/09/08/technology/google-salaries-gender-disparity .html.
35. Catherine D'Ignazio 和 Lauren F. Klein，《数据女权主义》（马萨诸塞州剑桥：麻省理工学院出版社，2020 年），65。
35. Catherine D’Ignazio and Lauren F. Klein, Data Feminism (Cambridge, MA: MIT Press, 2020), 65.
36. Sarah Hamid，“社区防御：Sarah T. Hamid 谈废除监狱技术”，《Logic Magazine》，2020 年 8 月 31 日，https://logicmag.io/care/community-defense-sarah-t-hamid-on-abolishing-carceral-technologies/。
36. Sarah Hamid, “Community Defense: Sarah T. Hamid on Abolishing Carceral Technologies,” Logic Magazine, August 31, 2020, https://logicmag.io/care/community-defense-sarah-t-hamid-on-abolishing-carceral -technologies/.
37. 鲁哈·本杰明 (Ruha Benjamin)，《科技之后的种族：新吉姆法典的废奴主义工具》 (英国剑桥；马萨诸塞州梅德福：Polity Press，2019 年)。
37. Ruha Benjamin, Race after Technology: Abolitionist Tools for the New Jim Code (Cambridge, UK; Medford, MA: Polity Press, 2019).
38. Zimmermann、Di Rosa 和 Kim，“技术无法解决算法不公正问题。”
38. Zimmermann, Di Rosa, and Kim, “Technology Can’t Fix Algorithmic Injustice.”
39. 齐默尔曼、迪罗莎和金。
39. Zimmermann, Di Rosa, and Kim.
40. Amy Kapczynski，《信息资本主义法则》，1460页。
40. Amy Kapczynski, “The Law of Informational Capitalism,” 1460.
41. 关于规范、法律、建筑和市场，请参见 Lawrence Lessig，《新芝加哥学派》，《法律研究杂志》第 27 卷，第 S2 期（1998 年）：第 661-91 页。
41. On norms, laws, architecture and markets, see, e.g., Lawrence Lessig, “The New Chicago School,” The Journal of Legal Studies 27, no. S2 (1998): 661–91.
所列页码与本书的印刷版相对应。您可以使用设备的搜索功能来查找文本中的特定术语。
Page numbers listed correspond to the print edition of this book. You can use your device’s search function to locate particular terms in the text.
埃森哲，288
Accenture, 288
阿加尔，乔恩，132–33
Agar, Jon, 132–33
英格兰和威尔士的贫困老人，（Booth），59
Aged Poor in England and Wales, The (Booth), 59
AI。参见人工智能算法。参见计算
AI. See artificial intelligence algorithms. See computational
技术；数据；数据分析；机器学习；神经网络；模式识别；回归
techniques; data; data analysis; machine learning; neural networks; pattern recognition; regression
艾伦，弗朗西斯，113–14
Allen, Frances, 113–14
阿尔斯通，菲利普·G.，233
Alston, Philip G., 233
亚马逊，4，215-16
Amazon, 4, 215–16
亚马逊的 Mechanical Turk，189，220
Amazon’s Mechanical Turk, 189, 220
美国经济协会，54
American Economic Association, 54
美国医学协会，95
American Medical Association, 95
《娱乐至死》（邮差），259–60
Amusing Ourselves to Death (Postman), 259–60
体质测量实验室（Galton），41，42
Anthropometric Laboratory (Galton), 41, 42
反垄断监管，296-299
antitrust regulation, 296–99
亚里士多德，51岁
Aristotle, 51
人工智能（AI），十二
artificial intelligence (AI), xii
当代意义，120
contemporary meaning of, 120
达特茅斯研讨会和，127–31，134，173
Dartmouth Workshop and, 127–31, 134, 173
数据驱动的方法，138–40
data-driven approach to, 138–40
数字计算机和 127–28, 129, 139–40
digital computers and, 127–28, 129, 139–40
道德与产品开发，233–34，245–46
ethics and product development in, 233–34, 245–46
专家系统，135–36，137–38，182–83
expert systems, 135–36, 137–38, 182–83
领域的基础，127–28
foundations in field of, 127–28
人类知识，131–33
human knowledge and, 131–33
人类劳动，219–21，274
human labor and, 219–21, 274
知识获取瓶颈，136–38，274
knowledge acquisition bottleneck, 136–38, 274
麦卡锡和，126–28, 133, 134, 135, 182
McCarthy and, 126–28, 133, 134, 135, 182
多层神经网络，177–79，185–86，188
multi-layer neural network and, 177–79, 185–86, 188
命名，126–27
naming of, 126–27
实际应用指标优化，172–74,194–95
real-world application metrics optimization and, 172–74, 194–95
重新定义机器学习，190–91
as redefined machine learning, 190–91
研究经费，133–35，182
research funding of, 133–35, 182
规则与数据，124–26，134，183
rules vs. data in, 124–26, 134, 183
象征性方法，123–24, 126–32, 134, 137–38, 175–79
symbolic approaches and, 123–24, 126–32, 134, 137–38, 175–79
象征性的反对，180, 183, 184, 185, 221
symbolic opposition and, 180, 183, 184, 185, 221
图灵和，121–23
Turing and, 121–23
另请参阅计算统计
See also computational statistics
AT&T 的贝尔实验室。参见贝尔实验室
AT&T’s Bell Labs. See Bell Labs
原子能委员会，115 n
Atomic Energy Commission, 115n
注意力经济
attention economy
算法优化的 UGC 以及 264–65
algorithmically optimized UGC and, 264–65
广播电视商业模式，258–60
broadcast television business model and, 258–60
互联网广告的起源，265–68
internet advertising, origins of, 265–68
优化参与度，269–70
optimizing engagement, 269–70
风险投资和 280–83
venture capital and, 280–83
万维网和，260–64
World Wide Web and, 260–64
自动化不平等（Eubanks），9
Automating Inequality (Eubanks), 9
反噬效应 274
backfire effect, 274
巴兰，保罗，162
Baran, Paul, 162
巴罗卡斯，索伦，251–52
Barocas, Solon, 251–52
贝叶斯，托马斯，109-10
Bayes, Thomas, 109–10
贝叶斯方法
Bayesian methods
贝叶斯规则，108–9，110–11
Bayes’ rule, 108–9, 110–11
工业数据分析，103，111-12
industrial data analysis and, 103, 111–12
统计，107–12, 116–17, 181
statistics, 107–12, 116–17, 181
钟形曲线，ix，26，31，32，38–39，73
bell curve, ix, 26, 31, 32, 38–39, 73
贝尔实验室，xii
Bell Labs, xii
现实世界系统中数据的积累，以及 206
accumulation of data in real-world systems and, 206
第二次世界大战后，120
after World War II, 120
通信数据和 118
communications data and, 118
数据科学和, 207, 208, 224
data science and, 207, 208, 224
集成建模和算法预测，188，193–94
ensemble modeling and algorithmic prediction, 188, 193–94
探索性数据分析，203–5，207
exploratory data analysis and, 203–5, 207
未来应用计算统计学的杰出人物，112
future luminaries of applied computational statistics and, 112
内核机器和 187
kernel machines and, 187
机器学习国际研究人员和184–85,187
machine learning international researchers and, 184–85, 187
机器学习和神经网络，171–72，183，185
machine learning and neural networks, 171–72, 183, 185
模式识别和，179–80
pattern recognition and, 179–80
另请参阅NSA
See also NSA
贝尔蒙特报告， 233，236，238–43，344n16
Belmont Report, The, 233, 236, 238–43, 344n16
Bengio，Yoshua，185
Bengio, Yoshua, 185
本杰明，鲁哈，7，251，304
Benjamin, Ruha, 7, 251, 304
边沁，杰里米，241
Bentham, Jeremy, 241
让·保罗·本塞克里，183–84
Benzécri, Jean-Paul, 183–84
伯内斯，爱德华，257–58，261，275–76
Bernays, Edward, 257–58, 261, 275–76
伯纳斯-李，蒂姆，260
Berners-Lee, Tim, 260
大数据
big data
5岁
age of, 5
分布式数据库平台，216-17
distributed database platform and, 216–17
电子数据处理转换，146，148-49
electronic data processing transformations, 146, 148–49
高维性，209
high-dimensionality of, 209
基础设施和工业智能伙伴关系，144–46，148–49，217
infrastructure and industrial-intelligence partnerships, 144–46, 148–49, 217
大规模私营行业数据收集，以及 142–43、155–56、166–67、224–25
mass private industry data collection and, 142–43, 155–56, 166–67, 224–25
技术整合，149–50
technological integration and, 149–50
技术驱动系统商业化，141–42
technology-driven systems commercialization, 141–42
另请参阅NSA；隐私和正义
See also NSA; privacy and justice
大桌子，214，217
Big Table, 214, 217
生物学, 43, 48, 57, 59, 199, 206
biology, 43, 48, 57, 59, 199, 206
生物特征识别实验室，44
Biometric Laboratory, 44
生物识别科学，50–53，57
biometric sciences, 50–53, 57
Biometrika（培生），83
Biometrika (Pearson), 83
美国黑人，54，56–57
Black Americans, 54, 56–57
“黑色星期五”，113
“Black Friday,” 113
布莱切利园，102–4, 106, 106 , 107, 111, 121, 139
Bletchley Park, 102–4, 106, 106, 107, 111, 121, 139
闪电扩张，281
blitzscaling, 281
体重指数，19
body mass index, 19
炸弹，103，105
bombes, 103, 105
查尔斯·布斯（Charles Booth） 59 岁，62 岁
Booth, Charles, 59, 62
博克，罗伯特，299
Bork, Robert, 299
安德鲁·博斯沃思，283
Bosworth, Andrew, 283
大卫·博特斯坦 210
Botstein, David, 210
Bottou，莱昂，187
Bottou, Léon, 187
布克，丹，14岁
Bouk, Dan, 14
博伊德，达纳，5，245，287
boyd, danah, 5, 245, 287
布兰德·斯图尔特，263
Brand, Stewart, 263
布兰代斯，路易斯，299
Brandeis, Louis, 299
Breiman，Leo，98，187，222–23
Breiman, Leo, 98, 187, 222–23
布林, 谢尔盖, 211, 212, 213, 214, 228, 268
Brin, Sergey, 211, 212, 213, 214, 228, 268
布朗，帕特里克，210
Brown, Patrick, 210
西蒙尼·布朗，14岁
Browne, Simone, 14
Brunsviga 机械计算器，64
Brunsviga mechanical calculator, 64
布莱恩·詹妮弗，226
Bryan, Jennifer, 226
布坎南，布鲁斯，136
Buchanan, Bruce, 136
Buolamwini，Joy，233
Buolamwini, Joy, 233
伯格，W.李，160
Burge, W. Lee, 160
伯克，科林，113
Burke, Colin, 113
剑桥分析公司，9，277
Cambridge Analytica, 9, 277
霍华德·坎皮恩 103
Campaigne, Howard, 103
资本主义，145
capitalism, 145
卡彭特诉美国，169
Carpenter v. United States, 169
卡森，约翰，73岁
Carson, John, 73
卡西利，安东尼奥，219–20
Casilli, Antonio, 219–20
种姓差异，52
caste difference, 52
凯瑟琳·考伊（Caughey），104岁
Caughey, Catherine, 104
因果关系，49，59–60，64–65
causation, 49, 59–60, 64–65
瑟夫，文特，5
Cerf, Vint, 5
约翰·钱伯斯，205–6，223
Chambers, John, 205–6, 223
中国情报机构，4
China Intelligence agencies, 4
乔杜里·鲁曼，288
Chowdhury, Rumman, 288
Chun, Wendy，5岁
Chun, Wendy, 5
阶级歧视，37，50，53
classism, 37, 50, 53
克利夫兰，威廉，196，207-8，223
Cleveland, William, 196, 207–8, 223
云计算，224–25
cloud-hosted computing, 224–25
认知科学，130
cognitive science, 130
科恩-科尔，杰米，130
Cohen-Cole, Jamie, 130
科恩，伯纳德，14岁
Cohn, Bernard, 14
冷战，144–46、196、197、218、276
Cold War, 144–46, 196, 197, 218, 276
巨人，103–4，106，139
Colossus, 103–4, 106, 139
《通信规范法》（1996 年），299
Communications Decency Act (1996), 299
康帕斯，251
COMPAS, 251
计算社会科学，243
computational social science, 243
计算统计
computational statistics
应用, 107, 112, 113, 116, 126, 140
applied, 107, 112, 113, 116, 126, 140
基于数据的预测重点，179–80
data-based predictions focus of, 179–80
工业智能资源和 185、216-19
industrial-intelligence resourcing and, 185, 216–19
NSA 和，112，168–69
NSA and, 112, 168–69
与数理统计相反，183–84,221
in opposition to mathematical statistics, 183–84, 221
计算技术 密码 贝叶斯分析 107–12, 116
computational techniques cryptographic Bayesian analysis and, 107–12, 116
数据挖掘，107，162，172-74，209-11，212-13，214-15，228-29
data mining, 107, 162, 172–74, 209–11, 212–13, 214–15, 228–29
机电计算设备和 103, 105
electromechanical computing devices and, 103, 105
历史机器计算，xii，44，52，64，75，105
historical machine calculation, xii, 44, 52, 64, 75, 105
打印表格，以及 46
printed tables and, 46
制表机，105
tabulators, 105
“密码组织推动的计算机进步” (Snyder)，115 n
“Computer Advances Pioneered by Cryptologic Organizations” (Snyder), 115n
计算机数据处理，103–4，113–14，117–18，144–46，148–49
computerized data processing, 103–4, 113–14, 117–18, 144–46, 148–49
计算机数据存储技术注意力经济和，258–60，269–70
computerized data storage technologies attention economy and, 258–60, 269–70
挑战，117–18
challenges of, 117–18
行业数据收集，以及 119
industry data collection and, 119
情报/工业/学术综合体，217
intelligence/industrial/academic complex and, 217
国家安全局发展支持，113–16，144–45，202
NSA development support of, 113–16, 144–45, 202
模式识别和139–40
pattern recognition and, 139–40
安全和劳动力成本，146
security and labor costs of, 146
开启，121–23
Turing on, 121–23
电脑，数码
computers, digital
人工智能和, 127–28, 129, 139–40
AI and, 127–28, 129, 139–40
EDP 应用转换，146，148-49
EDP applied transformations, 146, 148–49
逻辑和数学 vs. 数据和利润, 119
logic and math vs. data and profit, 119
数理统计和 223
mathematical statistics and, 223
神经网络和 177–80, 183, 185–91
neural networks and, 177–80, 183, 185–91
NSA 和 143–45
NSA and, 143–45
并行计算机和 186
parallel computers and, 186
收集、处理和存储数据的规模，145
scale in collecting, processing, and storing data, 145
计算机的起源，103–4，123–24
computers, origins of, 103–4, 123–24
布莱切利园研究人员和102–4
Bletchley Park researchers and, 102–4
编译器，133
compilers, 133
企业数据收集，119，142–43，145
corporate data collection and, 119, 142–43, 145
行业采用，142，144–45，148–50
industry adoption of, 142, 144–45, 148–50
NSA 技术发展，143–45
NSA technological development of, 143–45
编程语言，133
programming languages, 133
图灵和，121–23
Turing and, 121–23
“计算机器与智能”（图灵），121-23
“Computing Machinery and Intelligence” (Turing), 121–23
消费信贷行业，149，315 n 28
consumer credit industry, 149, 315n28
库克，蒂姆，286，292
Cook, Tim, 286, 292
卡尔文·柯立芝，275
Coolidge, Calvin, 275
公司权力，xii，305–7
corporate power, xii, 305–7
注意力经济/风险投资的影响，280–83
attention economy/venture capital consequences and, 280–83
数据赋能的算法系统重新调整社会秩序，6-10
data-empowered algorithmic systems realignment of social order and, 6–10
数据赋能技术分布，8，145-46
data-empowered technology distribution of, 8, 145–46
设计选择和误导，290–91
design choices and misdirection in, 290–91
道德原则和，243–48，286–90
ethical principles and, 243–48, 286–90
硬件到信息平台的技术转变，160 n
hardware to information platforms technology shifts and, 160n
用于数字问责，16-17
numerical accountability used for, 16–17
统计数据的起源，22-25
origins of statistics and, 22–25
personal data collection and interpretation history and, 22, 143, 150–60
与数据盈利能力的关系，285–86
relationship to data profitability and, 285–86
自律组织，293–94
self-regulatory organizations and, 293–94
国家权力作为监管者，294–96
state power as regulator and, 294–96
对个人的技术威胁，以及 13-14
technological threats to individuals and, 13–14
技术生态系统竞赛，291–93
technology ecosystem contests and, 291–93
另请参阅人民权力；国家权力
See also people power; state power
相关性
correlation
推理的危险，64–65
dangers of reasoning from, 64–65
Galton 和，40–43
Galton and, 40–43
Pearson 和 47, 49, 60
Pearson and, 47, 49, 60
物化，70
reification and, 70
Spearman 和，68–69，69
Spearman and, 68–69, 69
圣诞节和 63、64
Yule and, 63, 64
新冠肺炎，108–9，226，262
COVID-19, 108–9, 226, 262
克雷恩，马修，164，267，268
Crain, Matthew, 164, 267, 268
克劳福德，凯特，5岁
Crawford, Kate, 5
加密
cryptography
贝叶斯分析和107–12,116
Bayesian analysis and, 107–12, 116
数字计算缩放和，142–43，144
digital computation scaling and, 142–43, 144
模式识别和139–40
pattern recognition and, 139–40
权力动态和 105–7
power dynamics and, 105–7
储存和 114
storage and, 114
丹杰洛·亚当 269
D’Angelo, Adam, 269
丹齐格，库尔特，41，42
Danziger, Kurt, 41, 42
DARPA（国防高级研究计划局），134，138，170
DARPA (Defense Advanced Research Projects Agency), 134, 138, 170
“达特茅斯人工智能夏季研究项目” (McCarthy)，126–28
“Dartmouth Summer Research Project on Artificial Intelligence, The” (McCarthy), 126–28
达特茅斯研讨会，127–31，134，173
Dartmouth Workshop, 127–31, 134, 173
查尔斯·达尔文 33, 35, 36
Darwin, Charles, 33, 35, 36
数据（数据驱动的算法决策系统），x，xiii
data (data-driven algorithmic decision-making systems), x, xiii
反对激进主义，11
activism against, 11
进展，12-13
advances of, 12–13
算法优化的 UGC，264–65
algorithmically optimized UGC, 264–65
IRB 的应用伦理学，234–36，240–46
applied ethics of IRB and, 234–36, 240–46
贝尔蒙特报告和，238–40
Belmont report and, 238–40
内容审核，以及 274
content moderation and, 274
危险，7-10
dangers of, 7–10
定义数据科学领域，196–201
defining the field of data science, 196–201
虚假信息担忧，以及 10
disinformation concerns and, 10
早期网络技术和 142, 145, 156, 159
early networking technologies and, 142, 145, 156, 159
参与导向平台，270–73
engagement oriented platforms, 270–73
伦理与政治，xi，8–9，188，195，243–48
ethics and politics of, xi, 8–9, 188, 195, 243–48
未来潜力，306–7
future potential of, 306–7
历史和, 14–17, 152
history and, 14–17, 152
人类劳动，13–14, 146, 219–21, 274
human labor and, 13–14, 146, 219–21, 274
执法部门和 10
law enforcement and, 10
自由主义的政治观点, 6, 119, 153, 163, 164, 166
libertarian political views of, 6, 119, 153, 163, 164, 166
审计和分析方法，56
methods of auditing and analyzing, 56
政治说服架构和，274–78
political persuasion architecture and, 274–78
预测集成模型，187–91
predictive ensemble models, 187–91
预测系统和 179–80
predictive systems and, 179–80
真实世界数据指标优化，169–74，194–95
real-world data metrics optimization and, 169–74, 194–95
系统性不平等的再现，以及 7-8
reproduction of systemic inequalities and, 7–8
职责，12-14
responsibilities of, 12–14
秘密、专有、不透明处理，16，255–56
secret, proprietary, opaque processing in, 16, 255–56
伦理、公平和隐私方面的技术修复，249–54
technological fixes in ethics, fairness, and privacy, 249–54
万维网和，212–15，260–64
World Wide Web and, 212–15, 260–64
数据分析
data analysis
贝叶斯方法和 103, 107–12, 116–17, 181
Bayesian methods and, 103, 107–12, 116–17, 181
数据挖掘和，107，162，172-74，209-11，212-13，214-15，228-29
data mining and, 107, 162, 172–74, 209–11, 212–13, 214–15, 228–29
因子分析，70
factor analysis, 70
更多的统计数据，205–7，225–26
greater statistics and, 205–7, 225–26
历史数理统计，23–25，59–60，103
historical mathematical statistics, 23–25, 59–60, 103
数值差异的影响，52-53
implications in numerical differences, 52–53
工业规模统计数据，105，107
industrial-scale statistics, 105, 107
数学重点，理论统计，203
mathematically focused, theoretical statistics, 203
数理统计回归和39-40
mathematical statistics regression and, 39–40
NSA’s aggregate data analytical tools and, 168–69
纸质到计算机，98，103，118-19，202-8
paper to computer, 98, 103, 118–19, 202–8
序贯分析, 96
sequential analysis, 96
S 系统，207
S system for, 207
差异的统计分析，27–28,226–27
statistical analysis of differences and, 27–28, 226–27
统计和计算工具，204
statistical and computational tools for, 204
系统概况分析，73–75，149，150–51
systematic profile analysis and, 73–75, 149, 150–51
技术挑战，208
technological challenges in, 208
另请参阅数据科学；Pearson, Karl；Quetelet, Adolphe
See also data science; Pearson, Karl; Quetelet, Adolphe
数据库，148–49、152、155–60、213–14、216–17
databases, 148–49, 152, 155–60, 213–14, 216–17
数据收集
data collection
大数据收集规模，142–43、155–56、166–67、224–25
big data collection scale and, 142–43, 155–56, 166–67, 224–25
企业数据收集，119，141，142-43，145，161-62
corporate data collection, 119, 141, 142–43, 145, 161–62
企业数据挖掘，107，162，172-74
corporate data mining, 107, 162, 172–74
事实上的国家数据库和，152, 155–59
de facto national database and, 152, 155–59
数字计算机的历史规模，145
digital computers historical scale of, 145
制度化的数据使用/检索，111-12，113-18，119，148-49，152-53，160-61
institutionalized data usage/retrieval and, 111–12, 113–18, 119, 148–49, 152–53, 160–61
限制，169
limitations on, 169
正常化，145–46，159–60
normalization of, 145–46, 159–60
数值数据收集方法，23，41，43-44，73-75
numerical data collection methods, 23, 41, 43–44, 73–75
系统概况分析，73–75，149，150–51
systematic profile analysis and, 73–75, 149, 150–51
数据女权主义（D'Ignazio 和 Klein），219, 227, 303
Data Feminism (D’Ignazio and Klein), 219, 227, 303
数据新闻，227
data journalism, 227
数据挖掘，107，162，172-74，209-11，212-13，214-15，228-29
data mining, 107, 162, 172–74, 209–11, 212–13,214–15,228–29
遗传调查数据论文（Pearson），45
Data Paper for Heredity Investigations (Pearson), 45
数据处理、计算机化起源、103–4、113–14、117–18、144–46、148–49
data processing, computerized origins of, 103–4, 113–14, 117–18, 144–46, 148–49
“数据科学：扩大统计领域技术领域的行动计划” (克利夫兰)，196
“Data Science: An Action Plan for Expanding the Technical Areas of the Field of Statistics” (Cleveland), 196
数据科学
data science
广告和，270–73
advertising and, 270–73
IRB 流程的应用伦理学，234–36，240–46
applied ethics of IRB process and, 234–36, 240–46
贝尔实验室和，207，208，224
Bell Labs and, 207, 208, 224
克利夫兰，207–8
Cleveland and, 207–8
竞争性共同任务框架，193–94
competitive common task framework and, 193–94
数据驱动的种族主义，50
data-driven racism and, 50
数据挖掘，107，162，172-74，208-11，212-13，214-15，228-29
data mining, 107, 162, 172–74, 208–11, 212–13, 214–15, 228–29
定义角色，196–201
defining roles of, 196–201
开发新的架构，以及 214
development of new architectures and, 214
早期使用的关键算法，180
early use of key algorithms in, 180
道德和，228–29，233–36，238–40，241–46，247–48，249–54
ethics and, 228–29, 233–36, 238–40, 241–46, 247–48, 249–54
探索性数据分析，202-8
exploratory data analysis and, 202–8
历史，xi，48–49，52–53，224–26
history of, xi, 48–49, 52–53, 224–26
人类劳动，146，219–21，274
human labor and, 146, 219–21, 274
工业，197，200–201
industrial, 197, 200–201
工业智能资源和 185、216-19
industrial-intelligence resourcing and, 185, 216–19
工业规模机器学习，195
industrial-scale machine learning, 195
伪科学面相学，226–27
pseudoscientific physiognomy and, 226–27
实际应用指标优化，172–74,194–95
real-world application metrics optimization and, 172–74, 194–95
现实世界的统计数据，221–24
real-world statistics and, 221–24
研究和学术采用，225–28
research and academic adoption of, 225–28
统计数据与真实数据，199–201
statistics vs real-world data, 199–201
主题人性，6，238–40
subject humanity and, 6, 238–40
技术工具和，227–28
technology tools and, 227–28
万维网和，212-15
World Wide Web and, 212–15
另请参阅计算统计
See also computational statistics
戴维·南丁格尔，46 岁，83 岁
David, Florence Nightingale, 46, 83
深度学习，176，186，188，189，190
deep learning, 176, 186, 188, 189, 190
戴明，W.爱德华兹，96
Deming, W. Edwards, 96
丹德拉尔，136
DENDRAL, 136
阿兰·德罗西埃尔，39, 59, 67
Desrosières, Alain, 39, 59, 67
迪克，斯蒂芬妮，133岁
Dick, Stephanie, 133
迪迪埃·伊曼纽尔，14，74–75
Didier, Emmanuel, 14, 74–75
差异隐私，250
differential privacy, 250
数字计算，xii，44，52，64，75，105
digital computation, xii, 44, 52, 64, 75, 105
凯瑟琳·迪格纳齐奥，219、227、253、303
D’Ignazio, Catherine, 219, 227, 253, 303
埃琳娜·迪罗莎（Elena Di Rosa），290–91，304
Di Rosa, Elena, 290–91, 304
天意，25
Divine Providence, 25
多克托罗，科里，278
Doctorow, Cory, 278
唐纳利，凯文，25岁
Donnelly, Kevin, 25
多诺霍，大卫，194
Donoho, David, 194
多里奥特，乔治，279
Doriot, George, 279
DoubleClick，266，267
DoubleClick, 266, 267
烘干机，Theodora，90
Dryer, Theodora, 90
杜波依斯，网络，50–51, 53, 55–56
Du Bois, W. E. B., 50–51, 53, 55–56
杜布瓦松，多萝西，106
Du Boisson, Dorothy, 106
杜达，理查德，140
Duda, Richard, 140
杜姆比尔，埃德，141
Dumbill, Edd, 141
“二战期间”（Neyman），97
“During World War II” (Neyman), 97
辛西娅·德沃克，250
Dwork, Cynthia, 250
戴森，埃丝特，165
Dyson, Esther, 165
Eachus，Joseph，115 n
Eachus, Joseph, 115n
保罗·爱德华兹 149
Edwards, Paul, 149
埃夫隆·布拉德利，223
Efron, Bradley, 223
埃塞尔·埃尔德顿，44岁
Elderton, Ethel, 44
机电计算设备（炸弹），103，105
electromechanical computing devices (bombes), 103, 105
电子数据处理（EDP），148
electronic data processing (EDP), 148
实证技术，ix，24，52，93，125，194
empirical techniques, ix, 24, 52, 93, 125, 194
工程研究协会（ERA），144
Engineering Research Associates (ERA), 144
恩尼格玛密码机，103，105
Enigma machine, 103, 105
启蒙运动，21，22，28
Enlightenment, 21, 22, 28
埃尔文·萨姆 151
Ervin, Sam, 151
伦理算法（Kearns 和 Roth），253，289
Ethical Algorithm, The (Kearns and Roth), 253, 289
民族学证据，53
ethnological evidence, 53
弗吉尼亚州尤班克斯，8, 9, 251
Eubanks, Virginia, 8, 9, 251
欧几里得，93
Euclid, 93
优生学
eugenics
生物识别科学和 50–51
biometric sciences and, 50–51
Fisher 和，87–88
Fisher and, 87–88
Galton，37，70–71 n
Galton on, 37, 70–71n
历史意义，50
historical significance of, 50
历史观点, 38
historical view of, 38
智力测试，68–72
intelligence testing and, 68–72
自然选择，36–37
natural selection and, 36–37
正态曲线种族排名，38–39
normal curve racial ranking and, 38–39
Pearson 和，46–48
Pearson and, 46–48
优生学实验室，44
Eugenics Lab, 44
专家系统、人工智能和 135–36、137–38、182–83
expert systems, AI and, 135–36, 137–38, 182–83
探索性数据分析，202-8
exploratory data analysis, 202–8
广告模特，267
ad model, 267
人工智能伦理学家，7，11
AI ethicists, 7, 11
IRB 流程的应用伦理学，234–36，240–46
applied ethics of IRB process and, 234–36, 240–46
商业模式，4
business model of, 4
剑桥分析公司和 9
Cambridge Analytica and, 9
规模挑战，215–16
challenges of scale and, 215–16
情绪感染研究，234–36
emotional contagion study and, 234–36
公平流程，288
Fairness Flow, 288
Hadoop，216
Hadoop, 216
主持人和 274
moderators and, 274
说服架构以“相似受众”为特色，274–75
persuasion architecture feature “lookalike audiences,” 274–75
隐私和 160
privacy and, 160
风险资本主义和 282
venture capitalism and, 282
因子分析，70
factor analysis, 70
公平性/问责制/透明度问题
fairness/accountability/transparency concerns
算法系统和16–17、251–52、253、254–56
algorithmic systems and, 16–17, 251–52, 253, 254–56
贝尔蒙特原则和，241–43
Belmont principles and, 241–43
数据科学激励措施，6，234
data science incentives and, 6, 234
Facebook 公平流程和 288
Facebook Fairness Flow and, 288
运动，6–7，246–48
movement for, 6–7, 246–48
新的计算技术和劳动力成本，146，219-21，274
new computing technologies and labor costs, 146, 219–21, 274
新技术设计决策，149–50
new technology design decisions and, 149–50
优化目标，252–53
optimizing objectives and, 252–53
人民权力的吸收，7
people power co-optation of, 7
“负责任算法的原则”，第 11-12 页
“Principles for Accountable Algorithms” on, 11–12
量化公平，251–53，289–91
quantifying fairness and, 251–53, 289–91
技术方法，288–89
technical approaches to, 288–89
伦理、公平和隐私方面的技术修复，249–54
technological fixes in ethics, fairness, and privacy, 249–54
技术公司流程和，11，243-45
technology company processes and, 11, 243–45
瓦拉赫，3–4，250
Wallach on, 3–4, 250
范，詹妮弗，301
Fan, Jennifer, 301
法耶兹，乌萨马，208
Fayyad, Usama, 208
联邦贸易委员会（FTC），235
Federal Trade Commission (FTC), 235
费根鲍姆，爱德华，136
Feigenbaum, Edward, 136
爱德华·费尔顿 168
Felten, Edward, 168
First Census Optical Character Recognition System Conference, The, 171
第一个星期一（Goldhaber），261
First Monday (Goldhaber), 261
第一届世界种族代表大会（1911 年），50–51、53
First Universal Races Congress (1911), 50–51, 53
费舍尔，罗纳德，79，83–89，92，93，107
Fisher, Ronald, 79, 83–89, 92, 93, 107
费舍尔，RA，269
Fisher, R. A., 269
五眼联盟，106–7
Five Eyes, 106–7
美国食品药品管理局（FDA），95
Food and Drug Administration (FDA), 95
外国情报监视法庭（FISC），167
Foreign Intelligence Surveillance Court (FISC), 167
福柯，米歇尔，14，284
Foucault, Michel, 14, 284
第十四修正案，55
Fourteenth Amendment, 55
第四修正案，168
Fourth Amendment, 168
弗雷德里克森，乔治，56岁
Frederickson, George, 56
大卫·弗里德曼，65 岁，66 岁
Freedman, David, 65, 66
法国大革命，19
French Revolution, 19
弗里德曼，米尔顿，163
Friedman, Milton, 163
“从关联到因果关系：关于统计学历史的一些评论” (Freedman)，66
“From Association to Causation: Some Remarks on the History of Statistics” (Freedman), 66
高尔顿，弗朗西斯，40，42
Galton, Francis, 40, 42
数据相关性，40-43
data correlation and, 40–43
人类素质改善，36-37
on human quality improvement, 36–37
论天赋，70–71
on natural ability, 70–71
关于自然平等，37-38
on natural equality, 37–38
对民族和种族进行排名/分类，38–39，73
ranking/classifying peoples and races and, 38–39, 73
“个体差异的科学”，33、34、71
“science of individual differences” and, 33, 34, 71
统计回归和39-40
statistical regression and, 39–40
甘迪，奥斯卡，Jr.，5，164
Gandy, Oscar, Jr., 5, 164
高斯，卡尔，26岁
Gauss, Carl, 26
高斯曲线，ix
Gaussian curve, ix
GCHQ（政府通信总部）（英国），6，112，166
GCHQ (Government Communications Headquarters) (United Kingdom), 6, 112, 166
格布鲁、蒂姆尼特, 7, 233, 234, 246
Gebru, Timnit, 7, 233, 234, 246
戈登·盖柯 284
Gekko, Gordon, 284
通用数据保护条例（GDPR），297
General Data Protection Regulation (GDPR), 297
基因工程，243
genetic engineering, 243
德文密码，102–3
German cyphers, 102–3
加尼·拉伊德 276
Ghani, Rayid, 276
吉布森，威廉，8岁
Gibson, William, 8
金斯伯格，艾伦，196–97
Ginsberg, Allen, 196–97
吉特尔曼，丽莎，21岁
Gitelman, Lisa, 21
全球气候变化，58
global climate change, 58
全球反恐战争，217
Global War on Terror, 217
戈德哈伯，迈克尔，261，262，263
Goldhaber, Michael, 261, 262, 263
戈德华特，巴里，151
Goldwater, Barry, 151
好，杰克，112，139
Good, Jack, 112, 139
谷歌
广告和 263
ads and, 263
人工智能伦理学家，7，11，233–34，247，253
AI ethicists, 7, 11, 233–34, 247, 253
分析真实世界数据，4
analyzing real-world data, 4
大桌子，217
Big Table, 217
规模挑战，215–16
challenges of scale and, 215–16
伦理道德，246–47
ethics and, 246–47
方面，288
Facets, 288
MapReduce，216
MapReduce, 216
PageRank，214，263
PageRank, 214, 263
搜索引擎创建，214
search engine creation, 214
搜索功能以及 5–6、17、211、212
search feature and, 5–6, 17, 211, 212
监视广告和 267
surveillance advertising and, 267
训练神经网络，190
training neural nets and, 190
风险资本主义和 282
venture capitalism and, 282
假设分析工具 288
What-If Tool, 288
威廉·戈塞特，79–83, 86, 91–92
Gosset, William, 79–83, 86, 91–92
古尔德，斯蒂芬·杰伊，72岁
Gould, Stephen Jay, 72
科学语法（Pearson），90，91
Grammar of Science (Pearson), 90, 91
图形处理单元 (GPU)，190
graphics processing units (GPU), 190
玛丽·L·格雷，219–20
Gray, Mary L., 219–20
更多的统计数据，205–7，225–26
greater statistics, 205–7, 225–26
格里梅尔曼，詹姆斯，273
Grimmelmann, James, 273
“在 NSA 的计算机陪伴下成长（绝密 Umbra）”，116 n
“Growing Up with Computers at NSA (Top Secret Umbra),” 116n
格鲁斯，乔尔，197–98
Grus, Joel, 197–98
吉尼斯，爱德华，78岁
Guinness, Edward, 78
比尔·格尔利 279
Gurley, Bill, 279
盖恩，伊莎贝尔，172，185
Guyon, Isabelle, 172, 185
哈金，伊恩，19，22，28，31，32
Hacking, Ian, 19, 22, 28, 31, 32
海地革命，19
Haitian Revolution, 19
哈米德，莎拉T.，303–4
Hamid, Sarah T., 303–4
哈默巴赫，杰夫，196, 197, 216, 224, 269
Hammerbacher, Jeff, 196, 197, 216, 224, 269
哈特，莫里茨，251–52
Hardt, Moritz, 251–52
哈特，彼得，140
Hart, Peter, 140
林千喜男 184岁
Hayashi Chikio, 184
遗传天才（高尔顿），38，70–71 n
Hereditary Genius (Galton), 38, 70–71n
“遗传的天赋和性格” （高尔顿），36
“Hereditary Talent and Character” (Galton), 36
海克，亨特，129
Heyck, Hunter, 129
希克斯，马尔，104
Hicks, Mar, 104
HIPAA（1996），155
HIPAA (1996), 155
霍布斯，托马斯，28岁
Hobbes, Thomas, 28
霍德斯，玛莎，14岁
Hodes, Martha, 14
霍夫曼，弗雷德里克，54–55, 56, 57, 59, 76
Hoffman, Frederick, 54–55, 56, 57, 59, 76
Hollerith 打孔卡机，75，169
Hollerith punched card machines, 75, 169
homme moyen（普通男人），27岁
homme moyen (average man), 27
霍珀，格雷斯，133
Hopper, Grace, 133
Hotelling，Harold，96，97，203
Hotelling, Harold, 96, 97, 203
休姆，大卫，109–10
Hume, David, 109–10
赫特森，杰万，227
Hutson, Jevan, 227
Hwang, Tim, 278
Hwang, Tim, 278
假设检验显著性水平
hypothesis testing significance levels
食谱方法，93–96
cookbook approach, 93–96
Fisher 和，79，84–86
Fisher and, 79, 84–86
Gossett 和，79、81–83
Gossett and, 79, 81–83
Neyman 和，79、89、91–93
Neyman and, 79, 89, 91–93
IBM，7，11，288
IBM, 7, 11, 288
IBM 卡片处理机（制表机），105
IBM card-processing machines (tabulators), 105
伊戈·莎拉，14岁
Igo, Sarah, 14
移民，71，75
immigration, 71, 75
产学联合体，11，200，217
industrial-academic complex, 11, 200, 217
工业革命，13
Industrial Revolution, 13
不等式
inequality
科学，54–56
science of, 54–56
社会经济, 9–10, 50, 57
socioeconomic, 9–10, 50, 57
结构，7–8，253–54，302
structural, 7–8, 253–54, 302
信息处理系统，130
information processing systems, 130
机构审查委员会（IRB），234–36，240–46
institutional review board (IRB), 234–36, 240–46
智力测试，67–70，71–72，73
intelligence testing, 67–70, 71–72, 73
国际健康博览会（1884年），41
International Health Exhibition (1884), 41
互联网, 159, 162, 164–66, 224, 264–68, 299–301
internet, 159, 162, 164–66, 224, 264–68, 299–301
另请参阅万维网
See also World Wide Web
发明技术/新科学成果
inventive technology/new science outcomes
学术“俘获”和10-11、254-55
academic “capture” and, 10–11, 254–55
广告技术和说服架构，277–78
adtech and persuasion architectures and, 277–78
COMPAS 犯罪累犯算法，251
COMPAS algorithm of criminal recidivism and, 251
数据挖掘结果，228–29
data mining results and, 228–29
情绪感染研究，234–36
emotional contagion study and, 234–36
谷歌搜索和 5–6
Google search and, 5–6
大数据收集/实践分析，206-7
large data collection/practical analysis and, 206–7
世界各地劳动者的困惑，146，219–21，274
obfuscation of laborers worldwide, 146, 219–21, 274
“项目匹配”和人员类别，158–59
“Project Match” and categories of people, 158–59
凯特莱的普通男性，27岁
Quetelet’s average man, 27
生物学和科学，199
science of biology and, 199
算法系统的社会经济、性别和种族差异，253–54
socioeconomic, sexual, and racial disparities of algorithmic systems, 253–54
“战略主题清单”和9-10
“Strategic Subjects List” and, 9–10
技术决定论，14，306
technological determinism and, 14, 306
风险投资和新产品，279–83
venture capital and new products, 279–83
另请参阅大数据；优生学、生物特征识别科学；公平/问责/透明度问题；国家安全局；种族主义；科学种族主义；监控资本主义
See also big data; eugenics, biometric sciences; fairness/accountability/ transparency concerns; NSA; racism; scientific racism; surveillance capitalism
《英国贫困化变化原因调查》（Yule），60–64
“Investigation into the Causes of Changes in Pauperism in England, An” (Yule), 60–64
伊拉尼，莉莉，220，221
Irani, Lilly, 220, 221
IRB。参见机构审查委员会 Ireland, Eleanor, 104
IRB. See institutional review board Ireland, Eleanor, 104
艾萨克·迈克，292
Isaac, Mike, 292
以色列情报机构，4
Israel Intelligence agencies, 4
杰克曼，莫莉，235
Jackman, Molly, 235
Jacob，SM，44岁
Jacob, S. M., 44
雅各布斯，简，283
Jacobs, Jane, 283
珍妮薇，威廉，284–85
Janeway, William, 284–85
JD 洛克菲勒的标准石油公司，296
J. D. Rockefeller’s Standard Oil Company, 296
吉姆·克劳，55岁
Jim Crow, 55
乔丹，迈克尔，191，289
Jordan, Michael, 191, 289
卡内尔瓦，劳里，235
Kanerva, Lauri, 235
k-匿名，250
k-anonymity, 250
康德，伊曼纽尔，241
Kant, Immanuel, 241
卡普津斯基，艾米，294，305
Kapczynski, Amy, 294, 305
Kaplan，Lyric，284，295，296
Kaplan, Lyric, 284, 295, 296
迈克尔·卡恩斯 253, 289, 295
Kearns, Michael, 253, 289, 295
基福弗-哈里斯修正案（1962年），95
Kefauver-Harris Amendment (1962), 95
凯尼恩，大卫，103
Kenyon, David, 103
内核机器, 187
kernel machines, 187
关键绩效指标 (KPI), 194
key performance indicator (KPI), 194
可汗，莉娜，298，299
Khan, Lina, 298, 299
克莱因·劳伦, 219, 227, 253, 303
Klein, Lauren, 219, 227, 253, 303
知识获取瓶颈，人工智能，136–38，274
knowledge acquisition bottleneck, AI, 136–38, 274
数据库中的知识发现 (KDD), 209
Knowledge Discovery in Databases (KDD), 209
克兰兹伯格第一定律，3，12
Kranzberg’s First Law, 3, 12
所罗门·库尔巴克（Solomon Kullback），105，118
Kullback, Solomon, 105, 118
Kullback-Leibler 散度，117
Kullback-Leibler divergence, 117
库兹韦尔，雷，290
Kurzweil, Ray, 290
兰利，帕特，175，192
Langley, Pat, 175, 192
拉普拉斯，皮埃尔-西蒙，26 岁
Laplace, Pierre-Simon, 26
劳尔，乔什，149
Lauer, Josh, 149
错误法则，25–27、30、32、39
law of errors, 25–27, 30, 32, 39
克劳德·列维-斯特劳斯，125
Lévi-Strauss, Claude, 125
Lebacqz，Karen，242
Lebacqz, Karen, 242
Le Bui，Matthew，254
Le Bui, Matthew, 254
LeCun，Yann，172，185，187
LeCun, Yann, 172, 185, 187
莱德伯格，约书亚，136
Lederberg, Joshua, 136
李爱丽丝 44 岁 47 岁
Lee, Alice, 44, 47
李开复 282
Lee, Kai-Fu, 282
埃里希·莱曼（Erich L. Lehmann），92 岁
Lehmann, Erich L., 92
较小统计数字，205
lesser statistics, 205
李飞飞，188
Li, Fei-Fei, 188
李小昌 185
Li, Xiaochang, 185
莱特希尔，詹姆斯，134，182
Lighthill, James, 134, 182
LinkedIn，215-16
LinkedIn, 215–16
奥德·洛德（Audre Lorde），287
Lorde, Audre, 287
低权利环境, 人, 8, 159
low-rights environments, people in, 8, 159
马基雅维利，尼科洛，51岁
Machiavelli, Niccolò, 51
机器学习
machine learning
广告支持的 UGC 托管网站，264–65
Ad-supported UGC-hosting sites, 264–65
先进技术，214
advanced technologies and, 214
分析真实世界数据，3-4
analyzing real-world data and, 3–4
贝尔实验室和，171–72、183、184–85、187
Bell Labs and, 171–72, 183, 184–85, 187
共同任务框架，193–94
common task framework and, 193–94
数据挖掘和，107，162，172-74，209-11，212-13，214-15，228-29
data mining and, 107, 162, 172–74, 209–11, 212–13, 214–15, 228–29
数据科学学者，199–201
data science academics and, 199–201
人力劳动，219–21
human labor and, 219–21
工业情报联系，139–40
industrial-intelligence ties and, 139–40
工业规模机器学习，195
industrial-scale machine learning, 195
NSA 和，117，216-17
NSA and, 117, 216–17
数值预测和分类，175，191-92
numerical prediction and classification, 175, 191–92
优化广告，270–73
optimized advertising, 270–73
模式识别基础和 139–40, 175, 181, 183
pattern recognition fundamentals and, 139–40, 175, 181, 183
模式识别起源，179–80
pattern recognition origins, 179–80
平台的权力，169
power of platforms, 169
实用工程传统，181，183
practical engineering traditions of, 181, 183
预测集成模型和神经网络，187–91
predictive ensemble models and neural nets, 187–91
问题解决模拟和137-38
problem solving emulation and, 137–38
伪科学面相学，226–27
pseudoscientific physiognomy, 226–27
实际应用指标优化，172–74,194–95
real-world application metrics optimization and, 172–74, 194–95
重新定义，190–91
redefinition of, 190–91
赞助工业研究，以及10-11
sponsored industrial research and, 10–11
另请参阅人工智能
See also artificial intelligence
麦迪根，大卫，223
Madigan, David, 223
马哈拉诺比斯，Prasanta C.，35、51–53、67–68、74
Mahalanobis, Prasanta C., 35, 51–53, 67–68, 74
马洛克的机器，52
Mallock’s Machine, 52
Manes，Stephen，160 n
Manes, Stephen, 160n
曼海姆，卡尔，284，295，296
Manheim, Karl, 284, 295, 296
曼斯菲尔德修正案（1969年），134
Mansfield Amendment (1969), 134
MapReduce，214
MapReduce, 214
密码破译大师，103
master codebreakers, 103
数学严谨性
mathematical rigor
贝叶斯分析和 102, 103, 116
Bayesian analysis and, 102, 103, 116
好处，15
benefits of, 15
生物特征数据，51-52
biometrical data and, 51–52
数据科学与学术统计，222-23
data science and academic statistics, 222–23
工程和 113
engineering and, 113
数理统计领域和 97
mathematical statistics field and, 97
数理统计专业，98
mathematical statistics profession and, 98
皮尔逊和43
Pearson and, 43
Quetelet 的数据抽象和 27
Quetelet’s data abstraction and, 27
统计努力和 93–94
statistical effort and, 93–94
Tukey 和，203–4
Tukey and, 203–4
数理统计，x，xii
mathematical statistics, x, xii
替代统计模型，183–84
alternative statistical models, 183–84
Breiman 等，222–23
Breiman and, 222–23
历史数据驱动的分析, 23–25, 59–60, 103
historical data-driven analysis, 23–25, 59–60, 103
infrastructure, 44, 46, 64, 75
马哈拉诺比斯和, 35, 51–53, 74
Mahalanobis and, 35, 51–53, 74
Neyman 和，97–98
Neyman and, 97–98
数值问责方法，15-17
numerical accountability methods and, 15–17
数值数据收集方法，23，41，43-44，73-75
numerical data collection methods, 23, 41, 43–44, 73–75
数值数据方法论历史，21–24，38–40
numerical data methodological history, 21–24, 38–40
人民的量化，14–15，22–23
quantification of peoples, 14–15, 22–23
凯特勒、皮尔逊的愿景和 43
Quetelet, Pearson’s vision and, 43
对个人进行排名/分类，41-43
ranking/classifying individuals, 41–43
回归分析，39-40
regression analysis, 39–40
第二次世界大战及，96–98，124，125
World War II and, 96–98, 124, 125
毛泽东 75岁
Mao Zedong, 75
麦卡锡，约翰
McCarthy, John
人工智能和126–28、133、134、135、182
artificial intelligence and, 126–28, 133, 134, 135, 182
批评，131，132
criticism of, 131, 132
机器智能和 124
machine intelligence and, 124
麦卡锡，凯瑟琳，157
McCarthy, Kathleen, 157
麦克卢汉，马歇尔，273
McLuhan, Marshall, 273
麦克尼利，斯科特，160
McNealy, Scott, 160
商人艾米丽（Emily），14岁
Merchant, Emily, 14
桑迪普·默蒂亚（Sandeep Mertia），74岁
Mertia, Sandeep, 74
META（机器学习、道德、透明度和问责制），187，288
META (Machine Learning, Ethics, Transparency and Accountability), 187, 288
元数据，166，167，168
metadata, 166, 167, 168
梅特卡夫，雅各布，245，248
Metcalf, Jacob, 245, 248
米奇，唐纳德，112，137
Michie, Donald, 112, 137
微软，3，288
Microsoft, 3, 288
微目标定位，9，265–68，276–78
microtargeting, 9, 265–68, 276–78
MIDAS（斯坦福数据挖掘），211
MIDAS (Mining Data At Stanford), 211
军国主义，145
militarism, 145
军事。参见NSA；海军研究办公室
military. See NSA; Office of Naval Research
约翰·斯图尔特·穆勒，54，76，241
Mill, John Stuart, 54, 76, 241
米勒，亚瑟，151
Miller, Arthur, 151
米勒·凯利 76 岁
Miller, Kelly, 76
明斯基，马文，132，135
Minsky, Marvin, 132, 135
米切尔，玛格丽特，7，233，234，246，287
Mitchell, Margaret, 7, 233, 234, 246, 287
米切尔，汤姆，191，289
Mitchell, Tom, 191, 289
米特尔施塔特，布伦特，245
Mittelstadt, Brent, 245
穆迪，胡安妮塔，101
Moody, Juanita, 101
莫斯，伊曼纽尔，245
Moss, Emanuel, 245
穆罕默德·哈利勒·纪伯伦，14、57
Muhammad, Khalil Gibran, 14, 57
Muigai，Wangui，73，74
Muigai, Wangui, 73, 74
多层神经网络、人工智能和177–79、185–86、188
multi-layer neural network, AI and, 177–79, 185–86, 188
马斯克，伊隆，290
Musk, Elon, 290
MYCIN，136
MYCIN, 136
中村丽莎，3，8
Nakamura, Lisa, 3, 8
赤裸社会（帕卡德），150
Naked Society, The (Packard), 150
拿破仑帝国，19
Napoleonic Empire, 19
纳拉亚南，Arvind，249，251
Narayanan, Arvind, 249, 251
美国国家标准与技术研究院 (NIST)，170，171，173
National Institute of Standards and Technology (NIST), 170, 171, 173
美国国家科学基金会（NSF），98
National Science Foundation (NSF), 98
国家安全局 (NSA)。参见NSA (国家安全局)
National Security Agency (NSA). See NSA (National Security Agency)
自然选择，36-37
natural selection, 36–37
经济的本质（雅各布斯），283
Nature of Economies, The (Jacobs), 283
奈夫，吉娜，225–26，291
Neff, Gina, 225–26, 291
新布兰代斯主义反垄断监管，299
neo-Brandeisian antitrust regulation, 299
Netflix，192–93、194
Netflix, 192–93, 194
纽曼，杰瑞，280
Neuman, Jerry, 280
诺伊曼，亚当，281
Neumann, Adam, 281
神经网络，177–80，183，185–91
neural networks, 177–80, 183, 185–91
纽厄尔，艾伦，129–31，133
Newell, Allen, 129–31, 133
纽曼，马克斯，139–40
Newman, Max, 139–40
杰西·内曼，89–93, 97, 107
Neyman, Jerzy, 89–93, 97, 107
尼古拉斯，汤姆，280
Nicholas, Tom, 280
南丁格尔，佛罗伦萨，18，33–34，36，73
Nightingale, Florence, 18, 33–34, 36, 73
9/11，6，166
9/11, 6, 166
海伦·尼森鲍姆，5, 254, 290
Nissenbaum, Helen, 5, 254, 290
尼克斯，亚历山大，277
Nix, Alexander, 277
诺贝尔经济学奖，130
Nobel Prize in Economics, 130
贵族，萨菲亚，5–6, 254, 290
Noble, Safiya, 5–6, 254, 290
佩吉·努南，20岁
Noonan, Peggy, 20
正态曲线，ix，26，31，32，38–39，73
normal curve, ix, 26, 31, 32, 38–39, 73
诺齐克，罗伯特，163
Nozick, Robert, 163
第二次世界大战后的国家安全局（NSA），113-18
NSA (National Security Agency) after World War II, 113–18
汇总数据分析工具，168–69
aggregate data analytical tools and, 168–69
大数据之前，101
before big data, 101
冷战时期的国家资本主义和144-46
Cold War state-driven capitalism and, 144–46
communications surveillance capability expansion and, 6, 166–67
计算统计学，112，168–69
computational statistics at, 112, 168–69
计算机技术发展和 113–16, 143–45, 202
computer technological developments and, 113–16, 143–45, 202
数据挖掘和 214–15
data mining and, 214–15
数字计算机和 143–45
digital computers and, 143–45
分布式数据库平台，216-17
distributed database platform and, 216–17
早期数据处理和存储需求，113–16
early data processing and storage needs of, 113–16
工业规模贝叶斯分析，103，111-12
industrial scale Bayesian analysis and, 103, 111–12
制度文化，218–19
institutional culture of, 218–19
大型数据集和 202, 206, 208
large data sets and, 202, 206, 208
机器学习和，117，216-17
machine learning and, 117, 216–17
元数据与内容监控，167–69
metadata vs. content surveillance and, 167–69
论文到计算机数据分析，98，103，118-19，202-8
paper to computer data analysis, 98, 103, 118–19, 202–8
模式识别技术，139–40, 175, 179–80
pattern recognition technologies and, 139–40, 175, 179–80
另请参阅贝尔实验室
See also Bell Labs
零假设，81，85，89
null hypothesis, 81, 85, 89
数值问责，15-17，22，31，151-52，155-56，159-60
numerical accountability, 15–17, 22, 31, 151–52, 155–56, 159–60
萨姆·努恩 153
Nunn, Sam, 153
NVIDIA，190
NVIDIA, 190
奈·罗伯特 50 岁
Nye, Robert, 50
Ochigame，Rodrigo，255
Ochigame, Rodrigo, 255
奥康纳，凯文，266–67
O’Conner, Kevin, 266–67
海军研究办公室，96，97，98，117，134，214-15
Office of Naval Research, 96, 97, 98, 117, 134, 214–15
奥尼尔，凯茜，6，199，251，288
O’Neil, Cathy, 6, 199, 251, 288
“复仇行动”，第105–6页
“Operation Vengeance,” 105–6
蒂姆·奥赖利 264
O’Reilly, Tim, 264
物种起源（达尔文），36
Origin of Species (Darwin), 36
帕卡德·万斯，150
Packard, Vance, 150
佩奇，拉里，211，212，213，214，268
Page, Larry, 211, 212, 213, 214, 268
PageRank，214
PageRank, 214
佩珀特·西摩，135
Papert, Seymour, 135
帕斯夸莱，弗兰克，11，254
Pasquale, Frank, 11, 254
帕蒂尔，Dhanurjay“DJ”，197
Patil, Dhanurjay “DJ,” 197
爱国者法案（2001），166
PATRIOT Act (2001), 166
模式分类和场景分析（Duda 和 Hart），140
Pattern Classification and Scene Analysis (Duda and Hart), 140
模式识别，139–40, 175, 179–80, 181, 183, 224–25
pattern recognition, 139–40, 175, 179–80, 181, 183, 224–25
皮尔逊，埃贡，89，91，92
Pearson, Egon, 89, 91, 92
皮尔逊，卡尔，45岁
Pearson, Karl, 45
关于, 43, 107
about, 43, 107
相关性，47，49，60
on correlation, 47, 49, 60
数据收集方法，43-44
data collection methods, 43–44
数据和机器计算, 44, 46
data and machine calculations, 44, 46
优生社会主义和46-48
eugenic socialism and, 46–48
关于知识，91
on knowledge, 91
密封，51
Seal on, 51
社会和生物科学，48–49
on social and biological sciences, 48–49
关于斯皮尔曼的一般智力假设，72
on Spearman’s hypothesis of general intelligence, 72
系统的数据收集和分析，71，73
systematic data collection and analysis, 71, 73
佩恩，乔尼，127，130
Penn, Jonnie, 127, 130
人民权力，305–6
people power, 305–6
算法问责制，250–54
algorithmic accountability and, 250–54
算法决策系统的危险，10，152-53，155，158-60
algorithmic decision systems dangers to, 10, 152–53, 155, 158–60
集体行动宣传，301-2
collective action publicity, 301–2
数据赋能能力的影响，13-14
data-empowered capabilities effect upon, 13–14
用于数字问责，15-16
numerical accountability used for, 15–16
股东积极主义，302
shareholder activism as, 302
科技员工，302-3
tech employees and, 302–3
用户公众参与或脱离，304-5
user public engagement or disengagement and, 304–5
另请参阅公司权力；国家权力
See also corporate power; state power
中华人民共和国，75
People’s Republic of China, 75
感知器，177，179
Perceptron, 177, 179
佩雷蒂，乔纳，261
Peretti, Jonah, 261
菲利普斯，克里斯托弗，94岁
Phillips, Christopher, 94
面相学，226–27
physiognomy, 226–27
庇古，亚瑟，64–65
Pigou, Arthur, 64–65
乔治·波利亚，130
Pólya, George, 130
Poon, Martha, 149, 315附28
Poon, Martha, 149, 315n28
波特, 西奥多, 15, 32, 43, 95
Porter, Theodore, 15, 32, 43, 95
邮差，尼尔，259，273
Postman, Neil, 259, 273
贫困
poverty
布斯和，59
Booth and, 59
相关性和因果关系，64–65
correlation and causation, 64–65
多元回归和 58
multiple regression and, 58
贫困化物化，66–67
pauperism reification and, 66–67
国家政策和 58–59
state policies and, 58–59
圣诞节和 59–64, 65
Yule and, 59–64, 65
预测（判别）归纳模型，184
predictive (discriminative) models of induction, 184
“可信赖算法的原则”，11-12
“Principles for Accountable Algorithms,” 11–12
隐私法（1974年），155-56
Privacy Act (1974), 155–56
隐私杂志（史密斯），157
Privacy Journal (Smith), 157
隐私和正义
privacy and justice
贝尔蒙特报告和，238–43
Belmont report and, 238–43
大数据收集规模，142–43、155–56、166–67、224–25
big data collection scale and, 142–43, 155–56, 166–67, 224–25
大数据与社会观点的淡化，163–69
big data and emaciated social views of, 163–69
数据正义和，253–54
data justice and, 253–54
事实上的国家数据库收集，152，155-59
de facto national database collection, 152, 155–59
差异隐私，250
differential privacy, 250
歧视性把关，以及
discriminatory gatekeeping and, 153
互联网商业化，164–65
internet commercialization and, 164–65
法律/技术分析，6–7，163–65，166–69
legal/technological analysis and, 6–7, 163–65, 166–69
立法，154–56
legislation for, 154–56
个人电脑和 159–60
personal computers and, 159–60
隐私保护研究委员会和，143，155
Privacy Protection Study Commission and, 143, 155
隐私与信息自由流通，152–54，161–62
privacy vs. free circulation of information, 152–54, 161–62
隐私的复兴，150–53
revival of privacy, 150–53
伦理、公平和隐私方面的技术修复，249–54
technological fixes in ethics, fairness, and privacy, 249–54
信息的价值，152–54，160，161
value of information, 152–54, 160, 161
扎克伯格，160
Zuckerberg on, 160
隐私保护研究委员会，143，155
Privacy Protection Study Commission, 143, 155
“私人生活？不是我们的！”（Manes），160页
“Private Lives? Not Ours!” (Manes), 160n
宣传（伯内斯），258
Propaganda (Bernays), 258
珀塞尔，西奥多，247
Purcell, Theodore, 247
p 值，84，94
p values, 84, 94
凯特勒，阿道夫，九世
Quetelet, Adolphe, ix
约, 19–20, 73
about, 19–20, 73
“普通人”数据抽象，19，26-28
“average man” data abstraction and, 19, 26–28
数据期刊出版物和 25
data journal publications and, 25
关于犯罪的数据，29-30
on data regarding crime, 29–30
辨别人性的本质，28-29
on discerning the nature of humanity, 28–29
影响，32–34
influence of, 32–34
作为弗洛伦斯·南丁格尔的灵感来源，18
as inspiration for Florence Nightingale, 18
社会物理学，18–20，26–28，29–32
social physics and, 18–20, 26–28, 29–32
统计分析，24，25-26
statistical analysis and, 24, 25–26
奎兰，J.罗斯，137
Quinlan, J. Ross, 137
科技之后的竞赛（本杰明），304
Race After Technology (Benjamin), 304
种族定罪，57
racial criminalization, 57
种族科学，54-57
racial sciences, 54–57
另请参阅科学种族主义
See also scientific racism
科学种族主义。参见科学种族主义
racism, scientific. See scientific racism
种族主义
racism
优生学和 50, 51
eugenics and, 50, 51
谷歌搜索和 5–6、17
Google search and, 5–6, 17
侵犯隐私权，153
invasions of privacy and, 153
吉姆·克劳，55岁
Jim Crow, 55
人寿保险，55–57
life insurance and, 55–57
放贷，15-16
money lending and, 15–16
差异的统计分析，27–28,226–27
statistical analysis of differences and, 27–28, 226–27
塔斯基吉研究，236–40
Tuskegee study, 236–40
另请参阅公平/问责/透明度问题；隐私和正义；人类差异研究
See also fairness/accountability/ transparency concerns; privacy and justice; study of human differences
拉吉，Inioluwa Deborah，246
Raji, Inioluwa Deborah, 246
随机对照试验 (RCT)，86，95
randomized controlled trials (RCTs), 86, 95
里斯，米娜，96–97，117，203
Rees, Mina, 96–97, 117, 203
里根，普里西拉，164
Regan, Priscilla, 164
回归
regression
因果关系，60
causation and, 60
作为基础分析工具，66
as foundational analytic tool, 66
Galton 和，39–40
Galton and, 39–40
圣诞节和 58、59–60、61、63
Yule and, 58, 59–60, 61, 63
“遗传地位向平庸回归” (高尔顿)，40
“Regression Towards Mediocrity in Hereditary Stature” (Galton), 40
物化，ix，31，66–67，70
reification, ix, 31, 66–67, 70
“数据科学之路” (Grus), 197–98
“Road to Data Science, The” (Grus), 197–98
莎拉·罗伯茨，219–20
Roberts, Sarah, 219–20
洛克菲勒基金会，119
Rockefeller Foundation, 119
罗德，伊桑，276
Roeder, Ethan, 276
罗加韦，菲利普，105
Rogaway, Phillip, 105
罗森布拉特，弗兰克，177
Rosenblatt, Frank, 177
罗森塔尔，凯特琳，15岁
Rosenthal, Caitlin, 15
罗斯，亚伦，253，289，295
Roth, Aaron, 253, 289, 295
SABRE（半自动商业研究环境）系统，142
SABRE (Semi-Automatic Business Research Environment) system, 142
SAGE（半自动地面环境）系统，142
SAGE (Semi-Automatic Ground Environment) system, 142
Salganik，Matthew，235，277
Salganik, Matthew, 235, 277
Schapire，Rob，185
Schapire, Rob, 185
斯科尔曼，卡洛塔·费伊，256，258
Schoolman, Carlota Fay, 256, 258
迈克尔·施拉格，194
Schrage, Michael, 194
施瓦茨，杰克，138
Schwartz, Jack, 138
科学种族主义
scientific racism
优生学和38-39
eugenics and, 38–39
距离测量，52–53
measure of distance and, 52–53
伪科学面相学和机器学习，226–27
pseudoscientific physiognomy and machine learning and, 226–27
心理应用，71–73
psychological application and, 71–73
科学决策与政治决策，58–59
science-based vs. political policymaking and, 58–59
个体差异科学，32–33
science of individual differences and, 32–33
塔斯基吉研究，236–40
Tuskegee study and, 236–40
另请参阅优生学；公平/问责/透明度问题；隐私和正义；种族主义；人类差异研究
See also eugenics; fairness/account-ability/transparency concerns; privacy and justice; racism; study of human differences
海豹，布拉金德拉纳特，50–51, 53
Seal, Brajendranath, 50–51, 53
自律组织（SRO），293-94
self-regulatory organizations (SROs), 293–94
塞尔福里奇，奥利弗，131
Selfridge, Oliver, 131
序贯分析, 96
sequential analysis, 96
序贯决策理论，181
sequential decision theory, 181
塞拉，理查德，256，258–59
Serra, Richard, 256, 258–59
塞特勒·希夫朗，68 岁
Setler, Shivrang, 68
性别歧视
sexism
人工智能辩论和，131–32
artificial intelligence debates and, 131–32
布莱切利园和，104
Bletchley Park and, 104
Galton 和，37–38
Galton and, 37–38
谷歌搜索和 5–6、17
Google search and, 5–6, 17
侵犯隐私权，153
invasions of privacy and, 153
差异的统计分析，27–28,226–27
statistical analysis of differences and, 27–28, 226–27
另请参阅公平/问责/透明度问题；隐私和正义
See also fairness/accountability/ transparency concerns; privacy and justice
沙利兹，科斯玛，199
Shalizi, Cosma, 199
克劳德·香农（Claude Shannon），112, 120, 127–28, 133
Shannon, Claude, 112, 120, 127–28, 133
维塔利·什马蒂科夫，249
Shmatikov, Vitaly, 249
肖特，乔迪，164
Short, Jodi, 164
西蒙，赫伯特，129–31，133，178，183，257，258
Simon, Herbert, 129–31, 133, 178, 183, 257, 258
史密斯，CR，141
Smith, C. R., 141
史密斯，R.布莱尔，141
Smith, R. Blair, 141
史密斯，罗伯特·E.，156，157
Smith, Robert E., 156, 157
斯诺登，爱德华，6岁
Snowden, Edward, 6
斯奈德，塞缪尔·S.，115 n
Snyder, Samuel S., 115n
社会变革
social change
数据驱动的互联网和技术问题，5-12，306
data-driven internet and technology concerns and, 5–12, 306
数据的经济价值和 150–62
data economic value and, 150–62
优生学和 50
eugenics and, 50
未来，306-7
future and, 306–7
数据中介，x–xi，4
mediation by data, x–xi, 4
凯特勒的社会物理学和 18–20, 26–28, 29–32
Quetelet’s social physics and, 18–20, 26–28, 29–32
贝尔蒙特报告，233，236，238–43，344附16
The Belmont report, 233, 236, 238–43, 344n 16
风险投资和 279–83
venture capital and, 279–83
另见隐私和正义；社会权力关系
See also privacy and justice; societal power relationships
社会物理学，18–21，26–28，29–32
social physics, 18–21, 26–28, 29–32
社会权力关系
societal power relationships
历史理解，18-23
historical understanding of, 18–23
人民的量化和分类，14–17，38–39
quantification and classification of peoples and, 14–17, 38–39
转化统计分析，56–57，145
transformative statistical analysis and, 56–57, 145
另请参阅企业权力；人民权力；国家权力
See also corporate power; people power; state power
孙正义 281
Son, Masayoshi, 281
苏联加密系统，113
Soviet encryption system, 113
斯皮尔曼，查尔斯，15，68–70，69，71–72，171
Spearman, Charles, 15, 68–70, 69, 71–72, 171
斯台普顿，克莱尔，303
Stapleton, Claire, 303
斯塔克，卢克，227
Stark, Luke, 227
国家主导的资本主义，144，145
state-driven capitalism, 144, 145
国家权力，xii，xiii，305-7
state power, xii, xiii, 305–7
广告技术和说服架构的影响和，276–78
adtech and persuasion architectures effects and, 276–78
算法决策系统使能，10，158–60
algorithmic decision systems enabling of, 10, 158–60
企业权力的定义和规范，294–96
corporate power defined and regulated by, 294–96
数据赋能技术分布，8，145-46
data-empowered technology distribution of, 8, 145–46
用于数字问责，16-17
numerical accountability used for, 16–17
origins of statistics and, 22–25
个人数据收集和解释历史，22，143，150-60
personal data collection and interpretation history and, 22, 143, 150–60
重新制定反垄断监管，296-299
reconfigured antitrust regulation, 296–99
第 230 条重新评估，299–301
Section 230 reevaluation, 299–301
对个人的技术威胁，以及 13-14
technological threats to individuals and, 13–14
全球权力与密码学，105–7
worldwide power and cryptography, 105–7
另请参阅企业实力；人力实力
See also corporate power; people power
研究人员的统计方法（Fisher），85
Statistical Methods for Research Workers (Fisher), 85
伦敦统计学会，23
Statistical Society of London, 23
统计数据，来源，21–34
statistics, origins of, 21–34
应用计算统计学，107，112，113，116，126，140
applied computational statistics, 107, 112, 113, 116, 126, 140
作者隐藏了数据，82–83
author obscured data and, 82–83
贝叶斯方法，107–12,116–17,181
Bayesian methods, 107–12, 116–17, 181
因果关系，49
causation and, 49
对庸俗统计数据的批评，20-21
criticisms of vulgar statistics, 20–21
假设检验显著性水平，79, 81–83, 84–86, 89, 91–96
hypothesis testing significance levels and, 79, 81–83, 84–86, 89, 91–96
工业规模统计分析, 105, 107
industrial-scale statistical analysis, 105, 107
先天差异，76，226-27
innate differences and, 76, 226–27
原意，21-22
original meaning, 21–22
政策制定和 59–60
policymaking and, 59–60
改革愿景，18–19，73，75
reform visions and, 18–19, 73, 75
治国方略、经济和 22–25
statecraft, economies and, 22–25
变革权力，56-57
transformative power, 56–57
斯坦加特，阿尔玛，124–25
Steingart, Alma, 124–25
斯蒂格勒，斯蒂芬，63，109–10
Stigler, Stephen, 63, 109–10
禁止网络盗版法案（SOPA）（2012），293
Stop Online Piracy Act (SOPA) (2012), 293
“战略主题清单”，10
“Strategic Subjects List,” 10
“学生。”参见 Gosset, William
“Student.” See Gosset, William
学生 t 检验，81
Student’s t-test, 81
人类差异研究
study of human differences
数据驱动的种族主义，50，226–27
data-driven racisms in, 50, 226–27
Galton 和，33、37–38、41、71、73
Galton and, 33, 37–38, 41, 71, 73
制度化，43
institutionalization of, 43
智力和种族，53
intelligence and race, 53
数据的局限性，56，57
limitations of the data, 56, 57
马哈拉诺比斯和，52–53，68
Mahalanobis and, 52–53, 68
皮尔逊和，71
Pearson and, 71
凯特勒和，27，28
Quetelet and, 27, 28
社会等级制度和 76
social hierarchies and, 76
社会问题和 58
social issues and, 58
Spearman 和 71
Spearman and, 71
圣诞节和 58
Yule and, 58
Stuetzle，Walter，223
Stuetzle, Walter, 223
萨奇曼，露西，123，220
Suchman, Lucy, 123, 220
超级泵感（艾萨克），292
Super Pumped (Isaac), 292
支持向量机（SVM），185
support-vector machines (SVMs), 185
苏里·悉达多（Suri Siddharth），219–220
Suri, Siddharth, 219–20
监视资本主义
surveillance capitalism
互联网商业化，164–66
commercialization of the internet and, 164–66
通用数据保护条例 (GDPR) 以及 297
General Data Protection Regulation (GDPR) and, 297
政府利用商业数据，155–60
governments using business data for, 155–60
互联网广告微目标定位和 265–68
internet advertising microtargeting and, 265–68
互联网云计算和 224–25
internet cloud-hosted computing and, 224–25
NSA 通信能力扩展和 6, 166–67
NSA communications capability expansion and, 6, 166–67
政治说服架构和，274–78
political persuasion architecture and, 274–78
优先考虑经济效率而非人类价值，160
prioritized economic efficiency over human values, 160
隐私与信息自由流通，152–54，161–62
privacy vs. free circulation of information, 152–54, 161–62
社会秩序和权力不平等的重新调整，6-10
realignment of social order and power inequities in, 6–10
另请参阅NSA Sutton, Rich, 185
See also NSA Sutton, Rich, 185
斯威尼，拉塔尼亚，249
Sweeney, Latanya, 249
系统数据概况分析, 73–75, 145, 150–51
systematic data profile analysis, 73–75, 145, 150–51
驯服机会（黑客），32–33
Taming of Chance, The (Hacking), 32–33
泰勒，弗雷德里克，15岁
Taylor, Frederick, 15
技术决定论，14，306
technological determinism, 14, 306
撒切尔，玛格丽特，31岁
Thatcher, Margaret, 31
概率论，89
theory of probability, 89
瑞安·蒂布希拉尼（Ryan Tibshirani） 226
Tibshirani, Ryan, 226
蒂尔曼·雪莉，199
Tilghman, Shirley, 199
亚当·图兹（Adam Tooze） 75 岁
Tooze, Adam, 75
《证券真实性法》（1933 年），第 15-16 页
Truth in Securities Act (1933), 15–16
泽伊内普·图费克奇, 8, 255–56, 274, 275, 277
Tufekci, Zeynep, 8, 255–56, 274, 275, 277
图基，约翰
Tukey, John
数据分析，118-19
on data analysis, 118–19
exploratory data analysis and, 201–5
数学分析判断，98
on mathematical analytical judgment, 98
NSA 和大型数据集，202，206，208
NSA and large data sets, 202, 206, 208
作为一名叛逆的统计学家，222
as renegade statistician, 222
图灵和，112
Turing and, 112
图灵，艾伦，102，103，105，111，112，121–23，135
Turing, Alan, 102, 103, 105, 111, 112, 121–23, 135
图灵奖，114
Turing Award, 114
图灵机，121
Turing machine, 121
特纳，弗雷德，165
Turner, Fred, 165
图罗，约瑟夫，265
Turow, Joseph, 265
独角兽，261
unicorn, 261
英国
United Kingdom
GCHQ（政府通信总部），6，112，166
GCHQ (Government Communications Headquarters), 6, 112, 166
帝国衰落，35-36
imperial decline, 35–36
济贫法，58，65–66
Poor Laws, 58, 65–66
美国诉琼斯，169
United States v. Jones, 169
UNIVAC ， 142，146，147，169
UNIVAC, 142, 146, 147, 169
美国人口普查，22
U.S. Census, 22
美国人口普查局，96，169，170，173，250
U.S. Census Bureau, 96, 169, 170, 173, 250
美国宪法，22，55
U.S. Constitution, 22, 55
美国国防部，133–34
U.S. Department of Defense, 133–34
美国司法部，293
U.S. Department of Justice, 293
美国移民法（1924 年），71
U.S. Immigration Act (1924), 71
美国国家安全局 (NSA)。参见NSA
U.S. National Security Agency (NSA). See NSA
美国总统大选（2016），9
U.S. presidential election (2016), 9
美国公共卫生服务部，236–37
U.S. Public Health Service, 236–37
“美国公共卫生服务塔斯基吉梅毒研究”，236-40
“U.S. Public Health Service Syphilis Study at Tuskegee,” 236–40
用户生成内容 (UGC)，264–65
user-generated content (UGC), 264–65
瓦洛尔，香农，246
Vallor, Shannon, 246
瓦普尼克，弗拉基米尔，184，187
Vapnik, Vladimir, 184, 187
VC：美国历史（尼古拉斯），280
VC: An American History (Nicholas), 280
风险投资，279–83
venture capital, 279–83
维克多·维西 152
Veysey, Victor, 152
沃林斯基，克里斯，193
Volinksy, Chris, 193
瓦格纳，本，247–48
Wagner, Ben, 247–48
沃尔德·亚伯拉罕，181
Wald, Abraham, 181
瓦拉赫, 汉娜, 3, 4, 250
Wallach, Hanna, 3, 4, 250
沃纳梅克，约翰，269
Wanamaker, John, 269
华纳，马克，235
Warner, Mark, 235
数学毁灭武器（奥尼尔），6
Weapons of Math Destruction (O’Neil), 6
韦伯，詹姆斯，247
Weber, James, 247
韦尼蒙特，杰奎琳，14，23
Wernimont, Jacqueline, 14, 23
威斯汀，艾伦，153
Westin, Alan, 153
惠勒，斯坦顿，150
Wheeler, Stanton, 150
惠特克，梅雷迪斯，8，191，247，303
Whittaker, Meredith, 8, 191, 247, 303
威克姆，哈德利，226
Wickham, Hadley, 226
第二次世界大战，96，97，98，101，105–6，119，279
World War II, 96, 97, 98, 101, 105–6, 119, 279
万维网，211，212-15，260-64，265
World Wide Web, 211, 212–15, 260–64, 265
另请参阅互联网
See also internet
布料商同业公会，43，44
Worshipful Company of Drapers, 43, 44
吴，蒂姆，298
Wu, Tim, 298
雅虎，212，216
Yahoo, 212, 216
山本五十六 105–6
Yamamoto, Isoroku, 105–6
余斌, 221, 224
Yu, Bin, 221, 224
Yule，Udny，58，59–64，65–66，273
Yule, Udny, 58, 59–64, 65–66, 273
齐默尔曼，安妮特，290–91，304
Zimmermann, Annette, 290–91, 304
祖博夫，肖沙娜，220，255
Zuboff, Shoshana, 220, 255
扎克伯格，马克，9，160，268，269，285–86
Zuckerberg, Mark, 9, 160, 268, 269, 285–86
版权所有 © 2023 Chris Wiggins 和 Matthew L. Jones
Copyright © 2023 by Chris Wiggins & Matthew L. Jones
版权所有
All rights reserved
第一版
First Edition
有关复制选择许可的信息
For information about permission to reproduce selections
从这本书中，写信给 WW Norton & Company, Inc.，
from this book, write to Permissions, W. W. Norton & Company, Inc.,
第五大道 500 号，纽约，NY 10110
500 Fifth Avenue, New York, NY 10110
有关批量购买特别折扣的信息，
For information about special discounts for bulk purchases,
请联系 WW Norton 特约销售人员
please contact W. W. Norton Special Sales at
specialsales@wwnorton.com 或 800–233-4830
specialsales@wwnorton.com or 800–233-4830
夹克设计：David Drummond
Jacket design: David Drummond
书籍设计：Lisa Buckley
Book design by Lisa Buckley
生产经理：安娜·奥勒
Production manager: Anna Oler
国会图书馆出版品目数据可用
Library of Congress Cataloging-in-Publication Data is available
ISBN 978–1-324–00673-2
ISBN 978–1-324–00673-2
ISBN 978–1-324–00674-9(ebk.)
ISBN 978–1-324–00674-9(ebk.)
WW Norton & Company, Inc., 500 Fifth Avenue, 纽约, NY 10110
W. W. Norton & Company, Inc., 500 Fifth Avenue, New York, N.Y. 10110
WW Norton & Company Ltd.，伦敦卡莱尔街 15 号，W1D 3BS
W. W. Norton & Company Ltd., 15 Carlisle Street, London W1D 3BS